From e4dd50a53766c494ce4a0c465fccb876b02440b6 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:04:29 +0100 Subject: [PATCH 01/12] build(deps): pin lerobot to v0.7.0 --- pyproject.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/pyproject.toml b/pyproject.toml index 84edd1a84..ea356df5d 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -15,7 +15,7 @@ dependencies = [ "websockets>=15.0.1", "uvicorn>=0.24.0", "psutil>=5.9.0", - "lerobot[core_scripts,feetech,training] @ git+https://github.com/huggingface/lerobot.git@v0.6.1", + "lerobot[core_scripts,feetech,training] @ git+https://github.com/huggingface/lerobot.git@v0.7.0", # Windows-only: real DirectShow camera names for /available-cameras so the # frontend can match a camera to its browser deviceId (issues #12, #16). # Guarded by try/except at the call site, so its absence degrades gracefully. From 86b6fffe5a65ddde27a1650c97650dda7fd1f342 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:03:58 +0100 Subject: [PATCH 02/12] fix(calibrate): home the motors through LeRobot's serial bus LeRobot replaced the Feetech SDK bus with a rustypot-backed SerialMotorsBus (huggingface/lerobot#4672). Its _get_half_turn_homings() now takes the homing offset sign of each motor, so the homing step raised a TypeError before the range recording started. set_half_turn_homings() does the reset, the read and the offset writes in one call, and also clamps the offsets the Feetech register cannot hold (huggingface/lerobot#4940). The operating mode is set by name, since the OperatingMode enum is deprecated. rustypot serializes access to the port behind a mutex, so the "Port is in use" error the retries matched on no longer exists. --- lelab/calibrate.py | 33 +++++---------------------------- tests/test_calibrate.py | 19 +++++++++++++++++++ 2 files changed, 24 insertions(+), 28 deletions(-) diff --git a/lelab/calibrate.py b/lelab/calibrate.py index bc4e241d7..dd9fcefbb 100644 --- a/lelab/calibrate.py +++ b/lelab/calibrate.py @@ -28,7 +28,6 @@ from typing import Any, Literal from lerobot.motors import MotorCalibration -from lerobot.motors.feetech import OperatingMode from lerobot.robots import ( Robot, make_robot_from_config, @@ -333,19 +332,12 @@ def _step_homing(self): # Disable torque to allow manual movement during recording self.device.bus.disable_torque() - for motor in self.device.bus.motors: - self.device.bus.write("Operating_Mode", motor, OperatingMode.POSITION.value) + self.device.bus.set_operating_mode("position") - self.device.bus.reset_calibration() - actual_positions = self.device.bus.sync_read("Present_Position", normalize=False) - logger.info(f"Current positions for homing: {actual_positions}") - - self._homing_offsets = self.device.bus._get_half_turn_homings(actual_positions) + # Resets the calibration, then writes the offsets that center each present position. + self._homing_offsets = self.device.bus.set_half_turn_homings() logger.info(f"Calculated homing offsets: {self._homing_offsets}") - for motor, offset in self._homing_offsets.items(): - self.device.bus.write("Homing_Offset", motor, offset) - def _step_range_recording(self): """Record range of motion as the user moves all joints.""" logger.info("Starting range recording step") @@ -397,18 +389,7 @@ def _step_range_recording(self): # Record positions until user completes step while not self._step_complete.is_set() and not self.stop_calibration: try: - # Try reading positions with retry on port contention - positions = None - for attempt in range(3): # Try up to 3 times - try: - positions = self.device.bus.sync_read("Present_Position", normalize=False) - break # Success, exit retry loop - except Exception as read_error: - if "Port is in use" in str(read_error) and attempt < 2: - time.sleep(0.01) # Short delay before retry - continue - else: - raise read_error # Re-raise if not port contention or final attempt + positions = self.device.bus.sync_read("Present_Position", normalize=False) if positions: # Validate the readings - filter out invalid/zero values @@ -444,11 +425,7 @@ def _step_range_recording(self): except CalibrationDiscontinuityError: raise except Exception as e: - if "Port is in use" in str(e): - logger.debug(f"Port busy during position read: {e}") - else: - logger.warning(f"Error reading positions during recording: {e}") - # Increase sleep time on error to reduce port contention + logger.warning(f"Error reading positions during recording: {e}") time.sleep(0.2) if self.stop_calibration: diff --git a/tests/test_calibrate.py b/tests/test_calibrate.py index 760ef9383..158c04f8f 100644 --- a/tests/test_calibrate.py +++ b/tests/test_calibrate.py @@ -139,6 +139,25 @@ def test_get_status_returns_independent_recorded_ranges_snapshot() -> None: } +def test_step_homing_uses_the_serial_bus_api() -> None: + """Autospec checks every bus call against LeRobot's SerialMotorsBus signatures.""" + from unittest.mock import create_autospec + + from lelab.calibrate import CalibrationManager + from lerobot.motors import SerialMotorsBus + + bus = create_autospec(SerialMotorsBus, instance=True) + bus.set_half_turn_homings.return_value = {"shoulder_pan": -12} + mgr = CalibrationManager() + mgr.device = type("Device", (), {"bus": bus})() + + mgr._step_homing() + + bus.disable_torque.assert_called_once_with() + bus.set_operating_mode.assert_called_once_with("position") + assert mgr._homing_offsets == {"shoulder_pan": -12} + + def test_status_polling_does_not_wait_for_recording_bus_read() -> None: """A slow recording read does not block status polling.""" from lelab.calibrate import CalibrationManager From e4f4f48b2903e29f301583ecc0ed425d80854211 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:03:59 +0100 Subject: [PATCH 03/12] fix(devices): release the serial port through the bus The forced release reached into the Feetech SDK's port_handler, which SerialMotorsBus does not have, so it silently did nothing. The new bus closes the port in disconnect() even when disabling the torque fails, which is the guarantee the fallback needs. --- lelab/utils/devices.py | 15 ++++-------- tests/test_calibrate.py | 12 ++++----- tests/test_devices.py | 54 ++++++++++++++++++++++------------------- 3 files changed, 40 insertions(+), 41 deletions(-) diff --git a/lelab/utils/devices.py b/lelab/utils/devices.py index 59cda0635..6ba985814 100644 --- a/lelab/utils/devices.py +++ b/lelab/utils/devices.py @@ -10,7 +10,6 @@ import logging import threading -from contextlib import suppress from typing import Any # Recording, teleoperation, inference and replay all drive the follower: each @@ -35,17 +34,13 @@ def safe_disconnect_device(device: Any, logger: logging.Logger, context: str = " def _force_close_device_resources(device: Any, logger: logging.Logger) -> None: """Best-effort release for serial/camera resources after disconnect fails.""" bus = getattr(device, "bus", None) - port_handler = getattr(bus, "port_handler", None) - if port_handler is not None: - with suppress(Exception): - port_handler.clearPort() - with suppress(Exception): - port_handler.is_using = False + if bus is not None: + # SerialMotorsBus.disconnect() closes the port even when disabling the torque fails. try: - port_handler.closePort() - logger.info("Force-closed serial port after disconnect failure") + bus.disconnect() + logger.info("Released serial port after disconnect failure") except Exception as exc: - logger.warning("Failed to force-close serial port after disconnect failure: %s", exc) + logger.warning("Serial port released, but disabling the torque failed: %s", exc) cameras = getattr(device, "cameras", None) if isinstance(cameras, dict): diff --git a/tests/test_calibrate.py b/tests/test_calibrate.py index 158c04f8f..929c3a138 100644 --- a/tests/test_calibrate.py +++ b/tests/test_calibrate.py @@ -330,16 +330,16 @@ def test_cleanup_device_force_releases_and_clears_when_disconnect_fails() -> Non device handle — otherwise the COM port stays busy and blocks the next run.""" from lelab.calibrate import CalibrationManager - class PortHandler: + class Bus: def __init__(self) -> None: - self.closed = False + self.released = False - def closePort(self) -> None: # noqa: N802 - mirrors LeRobot port handler API - self.closed = True + def disconnect(self) -> None: + self.released = True class Device: def __init__(self) -> None: - self.bus = type("Bus", (), {"port_handler": PortHandler()})() + self.bus = Bus() def disconnect(self) -> None: raise RuntimeError("Failed to write 'Torque_Enable' on id_=6") @@ -350,5 +350,5 @@ def disconnect(self) -> None: mgr._cleanup_device() - assert device.bus.port_handler.closed is True # force-released despite failure + assert device.bus.released is True # force-released despite failure assert mgr.device is None # handle cleared so a new calibration can start diff --git a/tests/test_devices.py b/tests/test_devices.py index eab3e209b..b00e1c1ed 100644 --- a/tests/test_devices.py +++ b/tests/test_devices.py @@ -2,21 +2,32 @@ import logging +from lerobot.motors import Motor, MotorNormMode, SerialMotorsBus -def test_safe_disconnect_force_closes_serial_port_after_disconnect_failure() -> None: - from lelab.utils.devices import safe_disconnect_device - class PortHandler: - def __init__(self) -> None: - self.cleared = False - self.closed = False - self.is_using = True +class _UnresponsiveRustypotBus: + """Stands in for an open rustypot.Bus whose motors no longer answer.""" + + def __init__(self) -> None: + self.closed = False + + def set_torque(self, ids, enabled, retries=0): + return dict.fromkeys(ids, "no status packet") + + def close(self) -> None: + self.closed = True + - def clearPort(self) -> None: # noqa: N802 - mirrors LeRobot port handler API - self.cleared = True +def _connected_bus() -> SerialMotorsBus: + bus = SerialMotorsBus( + port="/dev/null", motors={"gripper": Motor(6, "sts3215", MotorNormMode.RANGE_0_100)} + ) + bus._bus = _UnresponsiveRustypotBus() + return bus - def closePort(self) -> None: # noqa: N802 - mirrors LeRobot port handler API - self.closed = True + +def test_safe_disconnect_releases_serial_port_after_disconnect_failure() -> None: + from lelab.utils.devices import safe_disconnect_device class Camera: def __init__(self) -> None: @@ -27,34 +38,27 @@ def disconnect(self) -> None: class Device: def __init__(self) -> None: - self.bus = type("Bus", (), {"port_handler": PortHandler()})() + self.bus = _connected_bus() self.cameras = {"cam": Camera()} def disconnect(self) -> None: - raise RuntimeError("Failed to write 'Torque_Enable' on id_=6") + raise RuntimeError("camera is not connected") device = Device() + port = device.bus._bus safe_disconnect_device(device, logging.getLogger(__name__)) - assert device.bus.port_handler.cleared is True - assert device.bus.port_handler.is_using is False - assert device.bus.port_handler.closed is True + assert port.closed is True # released even though disabling the torque failed + assert device.bus.is_connected is False assert device.cameras["cam"].disconnected is True def test_safe_disconnect_uses_normal_disconnect_when_it_succeeds() -> None: from lelab.utils.devices import safe_disconnect_device - class PortHandler: - def __init__(self) -> None: - self.closed = False - - def closePort(self) -> None: # noqa: N802 - mirrors LeRobot port handler API - self.closed = True - class Device: def __init__(self) -> None: - self.bus = type("Bus", (), {"port_handler": PortHandler()})() + self.bus = _connected_bus() self.disconnected = False def disconnect(self) -> None: @@ -64,4 +68,4 @@ def disconnect(self) -> None: safe_disconnect_device(device, logging.getLogger(__name__)) assert device.disconnected is True - assert device.bus.port_handler.closed is False + assert device.bus.is_connected is True # the fallback did not run From 36e1b40c8c6d2e9ebf69dc126ef9b1b6f22ddb9f Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:05:09 +0100 Subject: [PATCH 04/12] fix(hardware): write the calibration with the torque off LeRobot now disables the torque before writing a kept calibration file, so that Feetech motors store it (huggingface/lerobot#4867): with SerialMotorsBus, disabling the torque also opens the Feetech EEPROM lock. Recording, teleoperation and replay wrote the calibration right after connecting, assuming the torque was already off. --- lelab/record.py | 6 ++++-- lelab/replay.py | 2 ++ lelab/teleoperate.py | 5 ++++- 3 files changed, 10 insertions(+), 3 deletions(-) diff --git a/lelab/record.py b/lelab/record.py index 74044b7bb..af731cbb0 100644 --- a/lelab/record.py +++ b/lelab/record.py @@ -875,11 +875,13 @@ def record_with_web_events(cfg: RecordConfig, web_events: dict) -> LeRobotDatase logger.error(f"❌ TELEOP CONNECTION: Failed to connect teleoperator: {e}") raise - # Torque is still disabled at this point, which is what writing the - # calibration registers requires. + # Feetech motors only store calibration registers with the torque off, + # which also opens their EEPROM lock. logger.info("Writing calibration to motors...") + robot.bus.disable_torque() robot.bus.write_calibration(robot.calibration) if teleop is not None: + teleop.bus.disable_torque() teleop.bus.write_calibration(teleop.calibration) try: diff --git a/lelab/replay.py b/lelab/replay.py index f9d736424..2e3bf6d6d 100644 --- a/lelab/replay.py +++ b/lelab/replay.py @@ -88,6 +88,8 @@ def handle_start_replay(request: ReplayRequest) -> dict[str, Any]: f"Could not connect to the follower arm on {request.follower_port}. " "Make sure it's plugged in and powered on, then try again." ) from e + # Feetech motors only store calibration registers with the torque off. + robot.bus.disable_torque() robot.bus.write_calibration(robot.calibration) robot.configure() except Exception as e: diff --git a/lelab/teleoperate.py b/lelab/teleoperate.py index 2f9981851..581a3ab97 100644 --- a/lelab/teleoperate.py +++ b/lelab/teleoperate.py @@ -179,9 +179,12 @@ def handle_start_teleoperation(request: TeleoperateRequest, websocket_manager=No "Make sure it's plugged in and powered on, then try again." ) from e - # Write calibration to motors' memory + # Feetech motors only store calibration registers with the torque off, + # which also opens their EEPROM lock. logger.info("Writing calibration to motors...") + robot.bus.disable_torque() robot.bus.write_calibration(robot.calibration) + teleop_device.bus.disable_torque() teleop_device.bus.write_calibration(teleop_device.calibration) # Connect cameras and configure motors From bfb824f4238c031c375ca131f9bf7fd4347bcc3f Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:10:01 +0100 Subject: [PATCH 05/12] fix(train): track runs with Trackio, which replaced W&B in LeRobot LeRobot dropped its W&B integration for Trackio (huggingface/lerobot#4954), so every training run died at argument parsing on the `--wandb.enable` flag LeLab always sent. - Send `--trackio.enable`, `--trackio.project` and `--trackio.space_id`. Entity, notes, run id and mode have no Trackio counterpart. - Default the Space to a private `/lelab-trackio` for a signed-in user: it gives the run a dashboard link, and LeRobot refuses HF Jobs without one. Signed out, a local run stays in this machine's Trackio database. - Capture the Space URL from LeRobot's "Track this run -->" line. Jobs saved with a W&B link keep it. - Stop sending `disable_artifact`: Trackio now skips checkpoint artifacts by default (huggingface/lerobot#5123), where LeLab's default would have copied every checkpoint into the Space. - Drop the W&B install gate: Trackio ships with `lerobot[training]`. --- frontend/src/components/jobs/JobsSection.tsx | 2 +- .../training/WandbInstallDialog.tsx | 68 ---------------- .../training/config/AdvancedCard.tsx | 65 +++------------ .../training/config/EssentialsCard.tsx | 63 +++------------ frontend/src/components/training/types.ts | 11 +-- frontend/src/hooks/useInstallExtra.ts | 2 +- frontend/src/lib/jobsApi.ts | 13 ++- frontend/src/pages/Training.tsx | 19 ++--- lelab/jobs.py | 81 ++++++++++++------- lelab/server.py | 21 ----- lelab/train.py | 35 +++----- lelab/utils/system.py | 16 ---- tests/test_jobs.py | 64 +++++++++++++-- tests/test_train.py | 24 +++--- 14 files changed, 175 insertions(+), 309 deletions(-) delete mode 100644 frontend/src/components/training/WandbInstallDialog.tsx diff --git a/frontend/src/components/jobs/JobsSection.tsx b/frontend/src/components/jobs/JobsSection.tsx index 042446475..2e189eb30 100644 --- a/frontend/src/components/jobs/JobsSection.tsx +++ b/frontend/src/components/jobs/JobsSection.tsx @@ -104,7 +104,7 @@ const JobsSection: React.FC = () => { ...j, state: s.state, metrics: s.metrics, - wandb_run_url: s.wandb_run_url, + tracking_url: s.tracking_url, checkpoint_count: s.checkpoint_count, }; }); diff --git a/frontend/src/components/training/WandbInstallDialog.tsx b/frontend/src/components/training/WandbInstallDialog.tsx deleted file mode 100644 index 3ead2f07c..000000000 --- a/frontend/src/components/training/WandbInstallDialog.tsx +++ /dev/null @@ -1,68 +0,0 @@ -import React from "react"; -import { - Dialog, - DialogContent, - DialogDescription, - DialogHeader, - DialogTitle, -} from "@/components/ui/dialog"; -import { useInstallExtra } from "@/hooks/useInstallExtra"; -import { - InstallProgress, - InstallTitleIcon, - RestartInstructions, - installTitle, -} from "./InstallProgress"; - -interface Props { - open: boolean; - onOpenChange: (open: boolean) => void; - installHint: string; -} - -const WandbInstallDialog: React.FC = ({ open, onOpenChange, installHint }) => { - const install = useInstallExtra("system/wandb-extra", open); - - return ( - - - - - - {installTitle(install.state, "Weights & Biases Not Installed")} - - - Install the wandb package to enable W&B logging. - - - -
- - Enabling W&B logging requires the{" "} - - wandb - {" "} - package, which isn't installed in this environment. Install it - to log this run to W&B. - - } - doneDescription={} - /> -
- - - ); -}; - -export default WandbInstallDialog; diff --git a/frontend/src/components/training/config/AdvancedCard.tsx b/frontend/src/components/training/config/AdvancedCard.tsx index fee8d443f..70bd886fa 100644 --- a/frontend/src/components/training/config/AdvancedCard.tsx +++ b/frontend/src/components/training/config/AdvancedCard.tsx @@ -251,74 +251,33 @@ const AdvancedCard: React.FC = ({ config, updateConfig }) - {config.wandb_enable && ( + {config.trackio_enable && ( <>
- Weights & Biases + Trackio
-
-
- - - updateConfig('wandb_notes', e.target.value || undefined) - } - placeholder="Training run notes..." - className="bg-slate-900 border-slate-600 text-white rounded-lg" - /> -
-
- - -
-
- - updateConfig('wandb_disable_artifact', checked) - } - /> - +

+ Signed in, runs sync to a private lelab-trackio Space on your account by default. +

)} - {!config.wandb_enable && } + {!config.trackio_enable && } {/* Misc */}
diff --git a/frontend/src/components/training/config/EssentialsCard.tsx b/frontend/src/components/training/config/EssentialsCard.tsx index 0103668e2..ebb521ad5 100644 --- a/frontend/src/components/training/config/EssentialsCard.tsx +++ b/frontend/src/components/training/config/EssentialsCard.tsx @@ -1,4 +1,4 @@ -import React, { useState } from 'react'; +import React from 'react'; import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card'; import { Input } from '@/components/ui/input'; import { NumberInput } from '@/components/ui/number-input'; @@ -14,8 +14,6 @@ import { import { ConfigComponentProps } from '../types'; import DatasetCombobox from '@/components/replay/DatasetCombobox'; import { DatasetItem } from '@/lib/replayApi'; -import WandbInstallDialog from '../WandbInstallDialog'; -import { useApi } from '@/contexts/ApiContext'; interface EssentialsCardProps extends ConfigComponentProps { datasets: DatasetItem[]; @@ -23,35 +21,6 @@ interface EssentialsCardProps extends ConfigComponentProps { } const EssentialsCard: React.FC = ({ config, updateConfig, datasets, datasetsLoading }) => { - const { baseUrl, fetchWithHeaders } = useApi(); - const [wandbDialogOpen, setWandbDialogOpen] = useState(false); - const [wandbInstallHint, setWandbInstallHint] = useState('pip install wandb'); - - const handleWandbToggle = async (checked: boolean) => { - if (!checked) { - updateConfig('wandb_enable', false); - return; - } - // Check availability before flipping the switch on. If wandb isn't - // importable in this lelab process, surface the same install flow used - // for the training extra (accelerate) instead of letting the user start - // a run that will fail. - try { - const r = await fetchWithHeaders(`${baseUrl}/system/wandb-extra`); - const data: { available: boolean; install_hint: string } = await r.json(); - if (data.available) { - updateConfig('wandb_enable', true); - } else { - setWandbInstallHint(data.install_hint); - setWandbDialogOpen(true); - } - } catch { - // Backend unreachable — let the user proceed; training start will - // surface the real error if wandb is genuinely missing. - updateConfig('wandb_enable', true); - } - }; - return ( @@ -140,35 +109,29 @@ const EssentialsCard: React.FC = ({ config, updateConfig, d
updateConfig('trackio_enable', checked)} className="data-[state=checked]:bg-green-500" /> -
- - - {config.wandb_enable && ( + {config.trackio_enable && (
-
diff --git a/frontend/src/components/training/types.ts b/frontend/src/components/training/types.ts index cd4882ddd..9bcc4498b 100644 --- a/frontend/src/components/training/types.ts +++ b/frontend/src/components/training/types.ts @@ -21,13 +21,10 @@ export interface TrainingConfig { // Output configuration resume: boolean; - // Weights & Biases - wandb_enable: boolean; - wandb_project?: string; - wandb_entity?: string; - wandb_notes?: string; - wandb_mode?: string; - wandb_disable_artifact: boolean; + // Trackio + trackio_enable: boolean; + trackio_project?: string; + trackio_space_id?: string; // Policy-specific parameters policy_device?: string; diff --git a/frontend/src/hooks/useInstallExtra.ts b/frontend/src/hooks/useInstallExtra.ts index f53ba137b..a7567d3b0 100644 --- a/frontend/src/hooks/useInstallExtra.ts +++ b/frontend/src/hooks/useInstallExtra.ts @@ -26,7 +26,7 @@ export interface UseInstallExtraResult { } /** - * Drives the backend extra-install flow (`accelerate`, `wandb`, …). Seeds state + * Drives the backend extra-install flow (`accelerate`, policy extras, …). Seeds state * from `${endpointPrefix}/install-status`, polls while installing, and exposes * install/retry handlers. Pass `enabled=false` to gate seeding on dialog open. */ diff --git a/frontend/src/lib/jobsApi.ts b/frontend/src/lib/jobsApi.ts index 8cd82f70e..0cfbfeda3 100644 --- a/frontend/src/lib/jobsApi.ts +++ b/frontend/src/lib/jobsApi.ts @@ -36,12 +36,9 @@ export interface TrainingRequest { save_freq: number; save_checkpoint: boolean; resume: boolean; - wandb_enable: boolean; - wandb_project?: string; - wandb_entity?: string; - wandb_notes?: string; - wandb_mode?: string; - wandb_disable_artifact: boolean; + trackio_enable: boolean; + trackio_project?: string; + trackio_space_id?: string; policy_device?: string; policy_use_amp: boolean; optimizer_type?: string; @@ -79,7 +76,7 @@ export interface JobRecord { hf_flavor: string | null; hf_repo_id: string | null; hf_job_url: string | null; - wandb_run_url: string | null; + tracking_url: string | null; checkpoint_count: number; } @@ -89,7 +86,7 @@ export interface JobProgressSnapshot { id: string; state: JobState; metrics: TrainingMetrics; - wandb_run_url: string | null; + tracking_url: string | null; checkpoint_count: number; } diff --git a/frontend/src/pages/Training.tsx b/frontend/src/pages/Training.tsx index 2dd9a23f3..fafe4bcd1 100644 --- a/frontend/src/pages/Training.tsx +++ b/frontend/src/pages/Training.tsx @@ -80,12 +80,9 @@ function configToRequest(c: TrainingConfig): TrainingRequest { save_freq: c.save_freq, save_checkpoint: c.save_checkpoint, resume: c.resume, - wandb_enable: c.wandb_enable, - wandb_project: c.wandb_project, - wandb_entity: c.wandb_entity, - wandb_notes: c.wandb_notes, - wandb_mode: c.wandb_mode, - wandb_disable_artifact: c.wandb_disable_artifact, + trackio_enable: c.trackio_enable, + trackio_project: c.trackio_project, + trackio_space_id: c.trackio_space_id, policy_device: c.policy_device, policy_use_amp: c.policy_use_amp, optimizer_type: c.optimizer_type, @@ -131,9 +128,7 @@ const ConfigurationMode: React.FC = () => { save_freq: 1000, save_checkpoint: true, resume: false, - wandb_enable: false, - wandb_mode: "online", - wandb_disable_artifact: false, + trackio_enable: false, policy_use_amp: false, optimizer_type: "adam", use_policy_training_preset: true, @@ -586,14 +581,14 @@ const MonitoringMode: React.FC<{ jobId: string }> = ({ jobId }) => { View on Hub ↗ )} - {job.wandb_run_url && ( + {job.tracking_url && ( - View on W&B ↗ + View metrics ↗ )} diff --git a/lelab/jobs.py b/lelab/jobs.py index cca712970..2710381d9 100644 --- a/lelab/jobs.py +++ b/lelab/jobs.py @@ -36,7 +36,7 @@ from queue import Empty, Queue from typing import Literal, Protocol, runtime_checkable -from pydantic import BaseModel +from pydantic import AliasChoices, BaseModel, Field from .train import TrainingRequest @@ -88,8 +88,11 @@ class JobRecord(BaseModel): hf_flavor: str | None = None hf_repo_id: str | None = None hf_job_url: str | None = None - # Captured from training stdout the first time wandb prints the run URL. - wandb_run_url: str | None = None + # Captured from training stdout the first time Trackio prints the run's Space URL. + # Records persisted before the W&B to Trackio switch stored it as wandb_run_url. + tracking_url: str | None = Field( + default=None, validation_alias=AliasChoices("tracking_url", "wandb_run_url") + ) # Number of checkpoints currently visible (local: filesystem; cloud: # Hub repo). Filled in by JobRegistry.list/get; persisted as zero. checkpoint_count: int = 0 @@ -141,20 +144,20 @@ def stop(self) -> None: ... def is_running(self) -> bool: ... def returncode(self) -> int | None: ... def stream_log_lines(self) -> list[LogLine]: ... - def wandb_run_url(self) -> str | None: ... + def tracking_url(self) -> str | None: ... # tqdm progress: "Training: 1%|▏ | 125/10000 [02:02<2:36:10, 1.05step/s]" _TQDM_RE = re.compile(r"Training:\s*\d+%[^|]*\|[^|]*\|\s*(\d+)/(\d+)\s*\[(?:[\d:]+)<([\d:]+)") -# Wandb prints something like "wandb: 🚀 View run at https://wandb.ai///runs/" -# when it boots. We capture the first URL of that shape we see. -_WANDB_URL_RE = re.compile(r"https://wandb\.ai/[^\s/]+/[^\s/]+/runs/[A-Za-z0-9]+") +# LeRobot logs "Track this run --> https://huggingface.co/spaces//" when the run +# syncs to a Space; a run kept in the local Trackio database prints a `trackio show` command. +_TRACKING_URL_RE = re.compile(r"Track this run --> (https://\S+)") -def extract_wandb_run_url(line: str) -> str | None: - match = _WANDB_URL_RE.search(line) - return match.group(0) if match else None +def extract_tracking_url(line: str) -> str | None: + match = _TRACKING_URL_RE.search(line) + return match.group(1) if match else None def _parse_duration(s: str) -> float | None: @@ -230,7 +233,7 @@ def __init__( self._stop_event = threading.Event() self._log_file_path = log_file_path self._log_file = None # type: ignore[assignment] - self._wandb_run_url: str | None = None + self._tracking_url: str | None = None def _open_log_file(self) -> None: """Open the persistent log sink (one JSON line per consumed line). @@ -306,8 +309,8 @@ def stream_log_lines(self) -> list[LogLine]: pass return out - def wandb_run_url(self) -> str | None: - return self._wandb_run_url + def tracking_url(self) -> str | None: + return self._tracking_url # -- internals -- @@ -328,10 +331,10 @@ def _consume_lines(self, lines: Iterable[str]) -> None: if not stripped: continue parse_metrics_into(stripped, self._metrics) - if self._wandb_run_url is None: - url = extract_wandb_run_url(stripped) + if self._tracking_url is None: + url = extract_tracking_url(stripped) if url is not None: - self._wandb_run_url = url + self._tracking_url = url self._on_line(stripped) log_line = LogLine(timestamp=time.time(), message=stripped) if self._log_file is not None: @@ -402,7 +405,7 @@ def __init__( # Replay everything that's already on disk so the parser catches up # on metrics, then tail from the current EOF. self._tail_offset = 0 - self._wandb_run_url: str | None = None + self._tracking_url: str | None = None def start(self, job_id: str, config: TrainingRequest, output_dir: str) -> None: # Required by JobRunner Protocol but irrelevant here; the subprocess @@ -446,8 +449,8 @@ def stream_log_lines(self) -> list[LogLine]: def pid(self) -> int | None: return self._pid - def wandb_run_url(self) -> str | None: - return self._wandb_run_url + def tracking_url(self) -> str | None: + return self._tracking_url # -- internals -- @@ -476,10 +479,10 @@ def _tail_loop(self) -> None: except Exception: continue parse_metrics_into(log_line.message, self._metrics) - if self._wandb_run_url is None: - url = extract_wandb_run_url(log_line.message) + if self._tracking_url is None: + url = extract_tracking_url(log_line.message) if url is not None: - self._wandb_run_url = url + self._tracking_url = url if self._log_queue.qsize() >= 1000: with contextlib.suppress(Empty): self._log_queue.get_nowait() @@ -631,6 +634,23 @@ def _generate_job_id(policy_type: str, dataset_repo_id: str) -> str: return f"{policy_type}_{dataset_slug}_{timestamp}" +def _with_default_trackio_space(config: TrainingRequest) -> TrainingRequest: + """Sync a tracked run to the user's `lelab-trackio` Space when none is given. + + A Space is what gives the run a dashboard link, and LeRobot requires one for + HF Jobs, whose local Trackio database vanishes with the pod. Signed out, the + run stays in this machine's Trackio database. + """ + if not config.trackio_enable or config.trackio_space_id: + return config + from .utils.hf_auth import cached_whoami # lazy: keeps unit-test imports cheap + + info = cached_whoami() + if not info: + return config + return config.model_copy(update={"trackio_space_id": f"{info['name']}/lelab-trackio"}) + + def _job_dir(output_root: Path, job_id: str) -> Path: return output_root / job_id @@ -695,7 +715,7 @@ def __init__(self, output_root: Path) -> None: self._on_change: Callable[[], None] | None = None # Fired from the watchdog at ~1Hz with a compact snapshot of every - # running job (id, state, metrics, wandb url, checkpoint count) so + # running job (id, state, metrics, tracking url, checkpoint count) so # the dashboard keeps the progress bar live without refetching /jobs. self._on_progress: Callable[[builtins.list[dict]], None] | None = None @@ -818,6 +838,7 @@ def start(self, config: TrainingRequest, target: JobTarget | None = None) -> Job target = target or JobTarget() if target.runner == "hf_cloud" and not target.flavor: raise ValueError("flavor is required when runner is hf_cloud") + config = _with_default_trackio_space(config) with self._lock: # Local trainings are bounded by this machine's GPU/USB resources, @@ -1213,12 +1234,12 @@ def _tick(self) -> None: if runner is None or record is None: continue if runner.is_running(): - # Pull the wandb run URL once it appears in stdout. - if record.wandb_run_url is None: - url = runner.wandb_run_url() + # Pull the Trackio Space URL once it appears in stdout. + if record.tracking_url is None: + url = runner.tracking_url() if url is not None: with self._lock: - record.wandb_run_url = url + record.tracking_url = url self._persist(record, force=True) # Cloud jobs print their HF job id / page URL / model repo a few # seconds after start; capture them onto the record once parsed. @@ -1231,7 +1252,7 @@ def _tick(self) -> None: "id": record.id, "state": record.state, "metrics": record.metrics.model_dump(), - "wandb_run_url": record.wandb_run_url, + "tracking_url": record.tracking_url, "checkpoint_count": self._count_checkpoints(record), } ) @@ -1244,8 +1265,8 @@ def _tick(self) -> None: self._sync_cloud_ids(record, runner) rc = runner.returncode() with self._lock: - if record.wandb_run_url is None: - record.wandb_run_url = runner.wandb_run_url() + if record.tracking_url is None: + record.tracking_url = runner.tracking_url() record.state = "done" if rc == 0 else "failed" record.ended_at = time.time() record.exit_code = rc diff --git a/lelab/server.py b/lelab/server.py index dbb69ba68..33a7442f9 100644 --- a/lelab/server.py +++ b/lelab/server.py @@ -109,13 +109,10 @@ handle_get_cuda_status, handle_get_policy_extra, handle_get_training_extra, - handle_get_wandb_extra, handle_install_policy_extra, handle_install_policy_extra_status, handle_install_training_extra, handle_install_training_extra_status, - handle_install_wandb_extra, - handle_install_wandb_extra_status, warn_if_cuda_mismatch, ) @@ -937,24 +934,6 @@ def install_training_extra_status(): return handle_install_training_extra_status() -@app.get("/system/wandb-extra") -def get_wandb_extra(): - """Return whether the `wandb` package is importable in this lelab process.""" - return handle_get_wandb_extra() - - -@app.post("/system/wandb-extra/install") -def install_wandb_extra(): - """Spawn `pip install wandb` as a background subprocess. No-op if already running.""" - return handle_install_wandb_extra() - - -@app.get("/system/wandb-extra/install-status") -def install_wandb_extra_status(): - """Return current wandb install state plus any pending log lines (drained on read).""" - return handle_install_wandb_extra_status() - - @app.get("/system/policy-extra/{policy_type}") def get_policy_extra(policy_type: str): """Whether the optional LeRobot extra a policy needs (e.g. transformers for diff --git a/lelab/train.py b/lelab/train.py index 50419ad04..79a93ba1a 100644 --- a/lelab/train.py +++ b/lelab/train.py @@ -60,14 +60,11 @@ class TrainingRequest(BaseModel): resume: bool = False job_name: str | None = None - # Weights & Biases - wandb_enable: bool = False - wandb_project: str | None = None - wandb_entity: str | None = None - wandb_notes: str | None = None - wandb_run_id: str | None = None - wandb_mode: str | None = "online" - wandb_disable_artifact: bool = False + # Trackio. A run without a Space stays in this machine's Trackio database; + # JobRegistry.start defaults the Space for a signed-in user. + trackio_enable: bool = False + trackio_project: str | None = None + trackio_space_id: str | None = None # Environment / evaluation env_type: str | None = None @@ -201,20 +198,14 @@ def build_training_command( if request.job_name: cmd.extend(["--job_name", request.job_name]) - # W&B - cmd.extend(["--wandb.enable", "true" if request.wandb_enable else "false"]) - if request.wandb_enable: - if request.wandb_project: - cmd.extend(["--wandb.project", request.wandb_project]) - if request.wandb_entity: - cmd.extend(["--wandb.entity", request.wandb_entity]) - if request.wandb_notes: - cmd.extend(["--wandb.notes", request.wandb_notes]) - if request.wandb_run_id: - cmd.extend(["--wandb.run_id", request.wandb_run_id]) - if request.wandb_mode: - cmd.extend(["--wandb.mode", request.wandb_mode]) - cmd.extend(["--wandb.disable_artifact", "true" if request.wandb_disable_artifact else "false"]) + # Trackio + cmd.extend(["--trackio.enable", "true" if request.trackio_enable else "false"]) + if request.trackio_enable: + if request.trackio_project: + cmd.extend(["--trackio.project", request.trackio_project]) + if request.trackio_space_id: + # Private only applies when Trackio creates the Space; an existing one keeps its visibility. + cmd.extend(["--trackio.space_id", request.trackio_space_id, "--trackio.private", "true"]) # Env if request.env_type: diff --git a/lelab/utils/system.py b/lelab/utils/system.py index ed4023c7e..1693ddaba 100644 --- a/lelab/utils/system.py +++ b/lelab/utils/system.py @@ -33,9 +33,6 @@ TRAINING_AVAILABLE: bool = importlib.util.find_spec("accelerate") is not None TRAINING_INSTALL_HINT: str = "pip install accelerate" -WANDB_AVAILABLE: bool = importlib.util.find_spec("wandb") is not None -WANDB_INSTALL_HINT: str = "pip install wandb" - def _build_install_cmd(package: str) -> list[str]: """Pick the best installer for the running Python. @@ -159,7 +156,6 @@ def _drain_queue(self) -> None: training_install_manager = InstallManager("accelerate") -wandb_install_manager = InstallManager("wandb") def handle_get_training_extra() -> dict[str, Any]: @@ -174,18 +170,6 @@ def handle_install_training_extra_status() -> dict[str, Any]: return training_install_manager.get_status() -def handle_get_wandb_extra() -> dict[str, Any]: - return {"available": WANDB_AVAILABLE, "install_hint": WANDB_INSTALL_HINT} - - -def handle_install_wandb_extra() -> dict[str, Any]: - return wandb_install_manager.start() - - -def handle_install_wandb_extra_status() -> dict[str, Any]: - return wandb_install_manager.get_status() - - # --------------------------------------------------------------------------- # # Policy extras # --------------------------------------------------------------------------- # diff --git a/tests/test_jobs.py b/tests/test_jobs.py index 22f42d1a2..bfc68138d 100644 --- a/tests/test_jobs.py +++ b/tests/test_jobs.py @@ -21,18 +21,66 @@ import pytest -def test_extract_wandb_run_url_finds_canonical_url() -> None: - from lelab.jobs import extract_wandb_run_url +def test_extract_tracking_url_finds_the_space_url() -> None: + from lelab.jobs import extract_tracking_url - line = "wandb: \U0001f680 View run at https://wandb.ai/me/myproj/runs/abc123 trailing text" - assert extract_wandb_run_url(line) == "https://wandb.ai/me/myproj/runs/abc123" + line = "INFO 2026-11-05 15:03:26 io_utils.py:101 Track this run --> https://huggingface.co/spaces/me/lelab-trackio" + assert extract_tracking_url(line) == "https://huggingface.co/spaces/me/lelab-trackio" -def test_extract_wandb_run_url_returns_none_when_absent() -> None: - from lelab.jobs import extract_wandb_run_url +def test_extract_tracking_url_ignores_a_local_database_run() -> None: + from lelab.jobs import extract_tracking_url - assert extract_wandb_run_url("nothing here") is None - assert extract_wandb_run_url("https://example.com/runs/abc") is None + assert extract_tracking_url('Track this run --> trackio show --project "lerobot"') is None + assert extract_tracking_url("https://huggingface.co/spaces/me/other") is None + + +def test_job_record_reads_a_legacy_wandb_run_url() -> None: + from lelab.jobs import JobRecord + from lelab.train import TrainingRequest + + record = JobRecord.model_validate( + { + "id": "j", + "name": "n", + "state": "done", + "config": TrainingRequest(dataset_repo_id="x").model_dump(), + "output_dir": "/tmp/out", + "started_at": 0.0, + "wandb_run_url": "https://wandb.ai/me/p/runs/abc", + } + ) + assert record.tracking_url == "https://wandb.ai/me/p/runs/abc" + assert "wandb_run_url" not in record.model_dump() + + +@pytest.mark.parametrize( + ("whoami", "space_id", "expected"), + [ + ({"name": "me"}, None, "me/lelab-trackio"), + ({"name": "me"}, "team/runs", "team/runs"), # an explicit Space wins + (None, None, None), # signed out: the run stays in the local database + ], +) +def test_tracked_runs_default_to_the_users_trackio_space(monkeypatch, whoami, space_id, expected) -> None: + from lelab import jobs + from lelab.train import TrainingRequest + from lelab.utils import hf_auth + + monkeypatch.setattr(hf_auth, "cached_whoami", lambda: whoami) + config = TrainingRequest(dataset_repo_id="x", trackio_enable=True, trackio_space_id=space_id) + + assert jobs._with_default_trackio_space(config).trackio_space_id == expected + + +def test_untracked_runs_get_no_trackio_space(monkeypatch) -> None: + from lelab import jobs + from lelab.train import TrainingRequest + from lelab.utils import hf_auth + + monkeypatch.setattr(hf_auth, "cached_whoami", lambda: {"name": "me"}) + + assert jobs._with_default_trackio_space(TrainingRequest(dataset_repo_id="x")).trackio_space_id is None def test_parse_duration_handles_mm_ss_and_hh_mm_ss() -> None: diff --git a/tests/test_train.py b/tests/test_train.py index 3d99988f9..494bd324a 100644 --- a/tests/test_train.py +++ b/tests/test_train.py @@ -60,27 +60,27 @@ def test_optional_dataset_fields_only_present_when_set() -> None: assert cmd2[idx + 1 : idx + 4] == ["0", "1", "2"] -def test_wandb_block_only_serialized_when_enabled() -> None: +def test_trackio_block_only_serialized_when_enabled() -> None: from lelab.train import TrainingRequest, build_training_command - off = build_training_command(TrainingRequest(dataset_repo_id="x", wandb_enable=False), "/tmp/out") - assert _arg_value(off, "--wandb.enable") == "false" - assert "--wandb.project" not in off + off = build_training_command(TrainingRequest(dataset_repo_id="x"), "/tmp/out") + assert _arg_value(off, "--trackio.enable") == "false" + assert "--trackio.project" not in off + assert "--trackio.space_id" not in off on = build_training_command( TrainingRequest( dataset_repo_id="x", - wandb_enable=True, - wandb_project="proj", - wandb_entity="me", - wandb_run_id="abc", + trackio_enable=True, + trackio_project="proj", + trackio_space_id="me/lelab-trackio", ), "/tmp/out", ) - assert _arg_value(on, "--wandb.enable") == "true" - assert _arg_value(on, "--wandb.project") == "proj" - assert _arg_value(on, "--wandb.entity") == "me" - assert _arg_value(on, "--wandb.run_id") == "abc" + assert _arg_value(on, "--trackio.enable") == "true" + assert _arg_value(on, "--trackio.project") == "proj" + assert _arg_value(on, "--trackio.space_id") == "me/lelab-trackio" + assert _arg_value(on, "--trackio.private") == "true" def test_push_to_hub_emits_repo_id_only_when_enabled() -> None: From 5f68e2675ee19b042fb789fef1ddf77213497ea7 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:11:33 +0100 Subject: [PATCH 06/12] fix(train): make the AMP switch drive accelerate's mixed precision LeRobot's parallel training framework (huggingface/lerobot#4010) takes mixed precision from `--accelerator.mixed_precision` alone; `--policy.use_amp` is only read by evaluation and inference now, so the switch no longer changed how a run trained. Send fp16 on CUDA and on HF Jobs pods, the torch.autocast default that v0.6.1 trained with and that evaluation still runs under, and keep use_amp so the saved config still runs inference in AMP. --- lelab/train.py | 9 ++++++++- tests/test_train.py | 19 +++++++++++++++++++ 2 files changed, 27 insertions(+), 1 deletion(-) diff --git a/lelab/train.py b/lelab/train.py index 79a93ba1a..a370869e1 100644 --- a/lelab/train.py +++ b/lelab/train.py @@ -151,11 +151,18 @@ def build_training_command( # LeRobot defaults push_to_hub=True and demands --policy.repo_id when so. # submit_to_hf also resets policy.device so the pod picks its own device. is_cloud = job_target is not None and job_target.runner == "hf_cloud" + device = None if is_cloud else request.policy_device or auto_select_torch_device().type if not is_cloud: - cmd.extend(["--policy.device", request.policy_device or auto_select_torch_device().type]) + cmd.extend(["--policy.device", device]) cmd.extend(["--policy.push_to_hub", "true" if request.policy_push_to_hub else "false"]) if request.policy_push_to_hub and request.policy_repo_id: cmd.extend(["--policy.repo_id", request.policy_repo_id]) + # Training takes mixed precision from accelerate alone (huggingface/lerobot#4010); use_amp + # above stays for the saved config, which evaluation and inference read. fp16 matches the + # torch.autocast default they run under on CUDA, and works on pre-Ampere GPUs such as the + # T4. HF Jobs pods are GPUs. + if request.policy_use_amp and (is_cloud or device == "cuda"): + cmd.extend(["--accelerator.mixed_precision", "fp16"]) # GR00T-specific policy flags. These options only exist on the groot config, # so emit them only for --policy.type=groot; sending them to another policy diff --git a/tests/test_train.py b/tests/test_train.py index 494bd324a..6aa4fb90f 100644 --- a/tests/test_train.py +++ b/tests/test_train.py @@ -198,6 +198,25 @@ def test_local_run_keeps_explicit_device() -> None: assert _arg_value(cmd, "--policy.device") == "cpu" +@pytest.mark.parametrize( + ("device", "runner", "expected"), + [("cuda", "local", "fp16"), ("cpu", "local", None), ("mps", "local", None), (None, "hf_cloud", "fp16")], +) +def test_amp_trains_in_fp16_on_gpus(device, runner, expected) -> None: + from lelab.jobs import JobTarget + from lelab.train import TrainingRequest, build_training_command + + target = JobTarget(runner=runner, flavor="a10g-small") if runner == "hf_cloud" else None + req = TrainingRequest(dataset_repo_id="x", policy_device=device, policy_use_amp=True) + cmd = build_training_command(req, "/tmp/out", job_target=target) + + assert _arg_value(cmd, "--policy.use_amp") == "true" + if expected is None: + assert "--accelerator.mixed_precision" not in cmd + else: + assert _arg_value(cmd, "--accelerator.mixed_precision") == expected + + def test_groot_policy_flags_are_emitted_only_for_groot() -> None: from lelab.train import TrainingRequest, build_training_command From 0a2d2a771b86e54998e01fa7a54ef7979174e6ce Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:14:03 +0100 Subject: [PATCH 07/12] fix(system): check policy extras against LeRobot's version ranges The check probed one module per policy, which no longer matches LeRobot: the `wallx` extra dropped torchdiffeq (huggingface/lerobot#4077), so WALL-OSS stayed "missing" even once installed. A probe also passes an environment upgraded from an older LeRobot, where transformers 5.5 still imports though the extras now pin 5.22. Read what `lerobot[extra]` requires from LeRobot's own metadata, through the extras it is built from, and report each package that is missing or out of range. New extras need no hand-picked probe, and a test fails when LeRobot renames or drops an extra LeLab offers. --- .../components/training/PolicyExtraDialog.tsx | 8 +- lelab/utils/system.py | 104 ++++++++++++------ tests/test_utils_system.py | 72 +++++++++--- 3 files changed, 128 insertions(+), 56 deletions(-) diff --git a/frontend/src/components/training/PolicyExtraDialog.tsx b/frontend/src/components/training/PolicyExtraDialog.tsx index ff97982a2..7719548d9 100644 --- a/frontend/src/components/training/PolicyExtraDialog.tsx +++ b/frontend/src/components/training/PolicyExtraDialog.tsx @@ -18,7 +18,7 @@ interface Props { open: boolean; onOpenChange: (open: boolean) => void; policyType: string; - packageName: string; // the probed module, e.g. "transformers" + packageName: string; // the requirements still unmet, e.g. "transformers<5.23.0,>=5.22.0" installTarget: string; // e.g. "lerobot[smolvla]" installHint: string; // e.g. "pip install 'lerobot[smolvla]'" } @@ -63,11 +63,11 @@ const PolicyExtraDialog: React.FC = ({ idleTitle={title} idleDescription={ <> - Training a {policyType} policy needs the{" "} + Training a {policyType} policy needs{" "} {packageName}{" "} - package (installed via{" "} + (installed via{" "} {installTarget}), - which isn't in this environment yet. Install it to train this policy. + which this environment doesn't have yet. Install it to train this policy. } doneDescription={} diff --git a/lelab/utils/system.py b/lelab/utils/system.py index 1693ddaba..d45768bd9 100644 --- a/lelab/utils/system.py +++ b/lelab/utils/system.py @@ -21,8 +21,10 @@ import sys import threading import time +from importlib import metadata from typing import Any +from packaging.requirements import Requirement from pydantic import BaseModel logger = logging.getLogger(__name__) @@ -175,40 +177,73 @@ def handle_install_training_extra_status() -> dict[str, Any]: # --------------------------------------------------------------------------- # # Some LeRobot policies import an optional extra at construction time; training # (or inference) otherwise dies with a buried ImportError once the subprocess is -# already running. Map each such policy to the module we probe and the -# ``pip install lerobot[extra]`` target. Policies not listed (act, vqbet, tdmpc) -# need nothing extra. -POLICY_EXTRAS: dict[str, tuple[str, str]] = { - # policy_type: (probe_module, install_target) - "smolvla": ("transformers", "lerobot[smolvla]"), - "pi0": ("transformers", "lerobot[pi]"), - "pi0_fast": ("transformers", "lerobot[pi]"), - "pi05": ("transformers", "lerobot[pi]"), - "xvla": ("transformers", "lerobot[xvla]"), - "evo1": ("transformers", "lerobot[evo1]"), - "eo1": ("qwen_vl_utils", "lerobot[eo1]"), - "vla_jepa": ("qwen_vl_utils", "lerobot[vla_jepa]"), - "molmoact2": ("peft", "lerobot[molmoact2]"), - "wall_x": ("torchdiffeq", "lerobot[wallx]"), - "multi_task_dit": ("diffusers", "lerobot[multi_task_dit]"), - "fastwam": ("diffusers", "lerobot[fastwam]"), - "lingbot_va": ("diffusers", "lerobot[lingbot_va]"), - "diffusion": ("diffusers", "lerobot[diffusion]"), - # groot pulls in peft (LoRA), diffusers, timm, dm-tree and decord; peft is a - # reliable, cross-platform stand-in for "the lerobot[groot] extra is present". - "groot": ("peft", "lerobot[groot]"), +# already running. Map each such policy to its ``lerobot[extra]``; what the +# extra requires is read from lerobot's own metadata. Policies not listed (act, +# vqbet, tdmpc) need nothing extra. +POLICY_EXTRAS: dict[str, str] = { + "smolvla": "smolvla", + "pi0": "pi", + "pi0_fast": "pi", + "pi05": "pi", + "xvla": "xvla", + "evo1": "evo1", + "eo1": "eo1", + "vla_jepa": "vla_jepa", + "molmoact2": "molmoact2", + "wall_x": "wallx", + "multi_task_dit": "multi_task_dit", + "fastwam": "fastwam", + "lingbot_va": "lingbot_va", + "diffusion": "diffusion", + "groot": "groot", } + +def _unmet_requirements(extra: str) -> list[Requirement]: + """Requirements of ``lerobot[extra]`` this environment lacks or holds at a version out of range. + + A version check, not only an import probe: an environment upgraded from an older LeRobot + can still import a package the new extra pins to a different range. + """ + requires = [Requirement(r) for r in metadata.requires("lerobot") or []] + unmet: list[Requirement] = [] + pending, seen = [extra], set() + while pending: + name = pending.pop() + if name in seen: + continue + seen.add(name) + for req in requires: + # Skip core requirements, whose markers hold with or without an extra. + if ( + req.marker is None + or not req.marker.evaluate({"extra": name}) + or req.marker.evaluate({"extra": ""}) + ): + continue + if req.name == "lerobot": # an extra built from other extras + pending.extend(req.extras) + continue + try: + installed = metadata.version(req.name) + except metadata.PackageNotFoundError: + unmet.append(req) + continue + if not req.specifier.contains(installed, prereleases=True): + unmet.append(req) + return unmet + + # One install manager per install target (lerobot[smolvla] / lerobot[pi] / …), # created lazily so pi0 and pi0_fast share the lerobot[pi] install. _policy_install_managers: dict[str, InstallManager] = {} def _policy_install_manager(policy_type: str) -> InstallManager | None: - spec = POLICY_EXTRAS.get(policy_type) - if spec is None: + extra = POLICY_EXTRAS.get(policy_type) + if extra is None: return None - target = spec[1] + target = f"lerobot[{extra}]" mgr = _policy_install_managers.get(target) if mgr is None: mgr = InstallManager(target) @@ -217,14 +252,14 @@ def _policy_install_manager(policy_type: str) -> InstallManager | None: def handle_get_policy_extra(policy_type: str) -> dict[str, Any]: - """Whether the optional extra a policy needs is importable right now. + """Whether the optional extra a policy needs is installed right now. - Probed live (not cached at import) so a restart after installing is picked + Checked live (not cached at import) so a restart after installing is picked up. Policies that need nothing report ``available`` so the UI never blocks them. """ - spec = POLICY_EXTRAS.get(policy_type) - if spec is None: + extra = POLICY_EXTRAS.get(policy_type) + if extra is None: return { "policy_type": policy_type, "needs_extra": False, @@ -233,16 +268,13 @@ def handle_get_policy_extra(policy_type: str) -> dict[str, Any]: "install_target": "", "install_hint": "", } - probe, target = spec - try: - available = importlib.util.find_spec(probe) is not None - except (ImportError, ValueError): - available = False + target = f"lerobot[{extra}]" + unmet = _unmet_requirements(extra) return { "policy_type": policy_type, "needs_extra": True, - "available": available, - "package": probe, + "available": not unmet, + "package": ", ".join(f"{req.name}{req.specifier}" for req in unmet), "install_target": target, "install_hint": f"pip install '{target}'", } diff --git a/tests/test_utils_system.py b/tests/test_utils_system.py index a6ffa5a43..29ae1e0f0 100644 --- a/tests/test_utils_system.py +++ b/tests/test_utils_system.py @@ -18,6 +18,8 @@ import logging import sys +import pytest + def test_build_install_cmd_contains_pip_and_package() -> None: from lelab.utils.system import _build_install_cmd @@ -130,20 +132,69 @@ def test_cuda_status_endpoint_returns_expected_shape(client) -> None: def test_policy_extra_maps_policies_to_install_targets() -> None: - """smolvla/pi0/pi0_fast/diffusion map to the right probe module + lerobot[extra].""" from lelab.utils.system import handle_get_policy_extra smol = handle_get_policy_extra("smolvla") assert smol["needs_extra"] is True - assert smol["package"] == "transformers" assert smol["install_target"] == "lerobot[smolvla]" assert "lerobot[smolvla]" in smol["install_hint"] - # pi0 and pi0_fast share the lerobot[pi] extra; diffusion uses diffusers. + # pi0, pi0_fast and pi05 share the lerobot[pi] extra. assert handle_get_policy_extra("pi0")["install_target"] == "lerobot[pi]" assert handle_get_policy_extra("pi0_fast")["install_target"] == "lerobot[pi]" - assert handle_get_policy_extra("diffusion")["package"] == "diffusers" - assert handle_get_policy_extra("diffusion")["install_target"] == "lerobot[diffusion]" + assert handle_get_policy_extra("wall_x")["install_target"] == "lerobot[wallx]" + + +def test_every_policy_extra_exists_in_the_installed_lerobot() -> None: + """Catches an extra LeRobot renames or drops on the next version bump.""" + from importlib import metadata + + from packaging.utils import canonicalize_name + + from lelab.utils.system import POLICY_EXTRAS + + provided = {canonicalize_name(e) for e in metadata.metadata("lerobot").get_all("Provides-Extra")} + assert {e for e in POLICY_EXTRAS.values() if canonicalize_name(e) not in provided} == set() + + +def _fake_lerobot_requires(monkeypatch, installed: dict[str, str]) -> None: + from lelab.utils import system + + requires = [ + 'torchcodec<0.18.0; sys_platform == "darwin" or sys_platform != "darwin"', # core requirement + 'transformers<5.23.0,>=5.22.0; extra == "transformers-dep"', + 'num2words<0.6.0,>=0.5.14; extra == "smolvla"', + 'lerobot[transformers-dep]; extra == "smolvla"', + ] + + def version(name: str) -> str: + if name not in installed: + raise system.metadata.PackageNotFoundError(name) + return installed[name] + + monkeypatch.setattr(system.metadata, "requires", lambda _dist: requires) + monkeypatch.setattr(system.metadata, "version", version) + + +@pytest.mark.parametrize( + ("installed", "available", "package"), + [ + ({"transformers": "5.22.0", "num2words": "0.5.14"}, True, ""), + ({"transformers": "5.5.0", "num2words": "0.5.14"}, False, "transformers<5.23.0,>=5.22.0"), + ({"transformers": "5.22.0"}, False, "num2words<0.6.0,>=0.5.14"), + ], +) +def test_policy_extra_checks_versions_through_nested_extras( + monkeypatch, installed, available, package +) -> None: + """A missing core requirement (torchcodec here) is not the extra's to report.""" + from lelab.utils import system + + _fake_lerobot_requires(monkeypatch, installed) + status = system.handle_get_policy_extra("smolvla") + + assert status["available"] is available + assert status["package"] == package def test_policy_extra_core_policy_needs_nothing() -> None: @@ -155,17 +206,6 @@ def test_policy_extra_core_policy_needs_nothing() -> None: assert act["install_target"] == "" -def test_policy_extra_available_reflects_find_spec(monkeypatch) -> None: - import importlib.util - - from lelab.utils import system - - monkeypatch.setattr(importlib.util, "find_spec", lambda name: object()) - assert system.handle_get_policy_extra("smolvla")["available"] is True - monkeypatch.setattr(importlib.util, "find_spec", lambda name: None) - assert system.handle_get_policy_extra("smolvla")["available"] is False - - def test_policy_extra_install_is_noop_for_core_policy() -> None: from lelab.utils.system import handle_install_policy_extra, handle_install_policy_extra_status From d30885a9001c77af875bf7d46e745b38747e7e29 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:17:47 +0100 Subject: [PATCH 08/12] fix(rollout): pass the task to LeRobot as a quoted string draccus decodes a string flag as YAML, so a task like "step 1: grab the cube" reached the policy as a dict, "yes" as True, and "null" as None since the draccus bump in LeRobot main. A language-conditioned policy then ran on the wrong instruction without any error. JSON-quote the task, which draccus decodes back to the exact string. --- lelab/rollout.py | 3 ++- tests/test_rollout.py | 28 ++++++++++++++++++++++++++++ 2 files changed, 30 insertions(+), 1 deletion(-) diff --git a/lelab/rollout.py b/lelab/rollout.py index fed7dbabb..67b08a519 100644 --- a/lelab/rollout.py +++ b/lelab/rollout.py @@ -344,7 +344,8 @@ def handle_start_inference(request: InferenceRequest) -> dict[str, Any]: "--robot.type=so101_follower", f"--robot.port={request.follower_port}", f"--robot.id={follower_id}", - f"--task={request.task}", + # JSON-quoted: draccus decodes a string as YAML, so "step 1: grab it" became a dict. + f"--task={json.dumps(request.task)}", f"--duration={request.duration_s}", *_rollout_inference_args(policy_path), ] diff --git a/tests/test_rollout.py b/tests/test_rollout.py index a34196bc9..333768685 100644 --- a/tests/test_rollout.py +++ b/tests/test_rollout.py @@ -253,6 +253,34 @@ def boom(*args, **kwargs): assert rollout.inference_active is False +@pytest.mark.parametrize("task", ['put "it" in: the box', "step 1: grab the cube", "yes", "null", ""]) +def test_handle_start_inference_passes_the_task_verbatim(monkeypatch, tmp_path, task) -> None: + """draccus decodes a string flag as YAML, so the task must reach it quoted.""" + from dataclasses import dataclass + + import draccus + + from lelab import rollout + + monkeypatch.setattr(rollout, "setup_follower_calibration_file", lambda cfg: "robot_a") + monkeypatch.setattr(rollout, "_resolve_policy_path", lambda ref: str(tmp_path / "model")) + commands: list[list[str]] = [] + + def capture(cmd, **kwargs): + commands.append(cmd) + raise OSError("stop before spawning") + + monkeypatch.setattr(rollout.subprocess, "Popen", capture) + rollout.handle_start_inference(_stub_request().model_copy(update={"task": task})) + + @dataclass + class RolloutTask: + task: str = "" + + task_arg = next(arg for arg in commands[0] if arg.startswith("--task=")) + assert draccus.parse(RolloutTask, args=[task_arg]).task == task + + def test_classify_outcome_ok_warns_and_fails() -> None: from lelab.rollout import _classify_outcome From ee4c208767f0c65e3dc7bcb8b7ffb4457ecfc31f Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:17:47 +0100 Subject: [PATCH 09/12] fix(rollout): give setup time to stop before killing the process LeRobot now connects the arm only after the policy has loaded (huggingface/lerobot#4866), and a stop signal during setup is only noted until the control loop starts. Killing the process 5s after Stop could land in the middle of connecting the arm and leave it powered. Wait up to 60s while setup runs; the 5s grace still applies once the loop has started. --- lelab/rollout.py | 12 ++++++++++-- tests/test_rollout.py | 24 ++++++++++++++++++++++++ 2 files changed, 34 insertions(+), 2 deletions(-) diff --git a/lelab/rollout.py b/lelab/rollout.py index 67b08a519..b5b04c89d 100644 --- a/lelab/rollout.py +++ b/lelab/rollout.py @@ -75,6 +75,13 @@ def validate_rotations(cls, cameras: dict[str, dict[str, Any]]) -> dict[str, dic # UI can present a "rollout time" separate from the multi-second policy # load + bus connect + camera connect setup overhead. _ROLLOUT_START_MARKER = "Rollout setup complete" +# How long Stop waits for a graceful exit before killing. Once the loop runs, +# teardown (back to the start pose, then disconnect) takes about 4s. During +# setup LeRobot only notes the signal and keeps loading the policy, then +# connecting the arm (huggingface/lerobot#4866), before tearing down: killing it +# mid-connect would leave the arm powered. +_STOP_GRACE_S = 5 +_SETUP_STOP_GRACE_S = 60 def _pump_stdout(proc: subprocess.Popen, log_handle) -> None: @@ -418,13 +425,14 @@ def handle_stop_inference() -> dict[str, Any]: if not inference_active or _inference_proc is None: return {"success": False, "status_code": 409, "message": "No inference is active"} proc = _inference_proc + grace_s = _STOP_GRACE_S if _inference_rollout_started_at is not None else _SETUP_STOP_GRACE_S try: proc.terminate() try: - proc.wait(timeout=5) + proc.wait(timeout=grace_s) except subprocess.TimeoutExpired: - logger.warning("Inference did not exit in 5s; killing") + logger.warning("Inference did not exit in %ss; killing", grace_s) proc.kill() proc.wait() except Exception as exc: diff --git a/tests/test_rollout.py b/tests/test_rollout.py index 333768685..c533f11ce 100644 --- a/tests/test_rollout.py +++ b/tests/test_rollout.py @@ -281,6 +281,30 @@ class RolloutTask: assert draccus.parse(RolloutTask, args=[task_arg]).task == task +@pytest.mark.parametrize( + ("rollout_started_at", "grace"), [(None, "_SETUP_STOP_GRACE_S"), (123.0, "_STOP_GRACE_S")] +) +def test_stop_gives_setup_longer_to_exit(monkeypatch, rollout_started_at, grace) -> None: + """Killed mid-setup, LeRobot may be connecting the arm and leave it powered.""" + from lelab import rollout + + waits: list[float] = [] + + class Proc: + def terminate(self) -> None: + pass + + def wait(self, timeout=None) -> None: + waits.append(timeout) + + monkeypatch.setattr(rollout, "inference_active", True) + monkeypatch.setattr(rollout, "_inference_proc", Proc()) + monkeypatch.setattr(rollout, "_inference_rollout_started_at", rollout_started_at) + + assert rollout.handle_stop_inference()["success"] is True + assert waits == [getattr(rollout, grace)] + + def test_classify_outcome_ok_warns_and_fails() -> None: from lelab.rollout import _classify_outcome From 67f5ebb0fe04cefaba7ea09b66d906e0df56894f Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:21:20 +0100 Subject: [PATCH 10/12] test: run the argv LeLab builds through LeRobot's own parsers The argv-shape tests all passed while every training run died on the removed `--wandb.enable` flag. Parse the real training argv (local, HF Jobs, and GR00T's own flags) and the real rollout argv with the installed LeRobot, and check every policy the training form offers is a policy type LeRobot registers, so the next rename fails here instead of on a user's run. --- tests/test_lerobot_cli.py | 132 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 132 insertions(+) create mode 100644 tests/test_lerobot_cli.py diff --git a/tests/test_lerobot_cli.py b/tests/test_lerobot_cli.py new file mode 100644 index 000000000..853cdf058 --- /dev/null +++ b/tests/test_lerobot_cli.py @@ -0,0 +1,132 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""The argv LeLab builds, run through the installed LeRobot's own CLI parsers. + +The argv-shape tests elsewhere cannot see a flag LeRobot renamed or dropped: +they all passed while every training run died on `--wandb.enable`. +""" + +# No `from __future__ import annotations`: parser.wrap reads the config class off the annotation. +import re +import sys +from pathlib import Path + +import pytest + +import lerobot.scripts.lerobot_train # noqa: F401 registers the policy configs, as the CLI does +from lerobot.configs import parser +from lerobot.configs.policies import PreTrainedConfig +from lerobot.configs.train import TrainPipelineConfig + +_ESSENTIALS_CARD = Path(__file__).parent.parent / "frontend/src/components/training/config/EssentialsCard.tsx" + + +@parser.wrap() +def _parse_training(cfg: TrainPipelineConfig) -> TrainPipelineConfig: + return cfg + + +def test_every_policy_the_ui_offers_is_a_lerobot_policy_type() -> None: + offered = set(re.findall(r'', _ESSENTIALS_CARD.read_text())) + assert offered, "policy list not found in EssentialsCard.tsx" + assert offered - set(PreTrainedConfig.get_known_choices()) == set() + + +@pytest.mark.parametrize( + ("policy_type", "runner"), [("act", "local"), ("act", "hf_cloud"), ("groot", "local")] +) +def test_training_argv_parses(monkeypatch, policy_type, runner) -> None: + from lelab.jobs import JobTarget + from lelab.train import TrainingRequest, build_training_command + + request = TrainingRequest( + dataset_repo_id="me/dataset", + policy_type=policy_type, + policy_device="cpu", + policy_use_amp=True, + trackio_enable=True, + trackio_project="lelab", + trackio_space_id="me/lelab-trackio", + optimizer_lr=1e-4, + **( + {"policy_embodiment_tag": "new_embodiment", "policy_relative_exclude_joints": ["gripper"]} + if policy_type == "groot" + else {} + ), + ) + target = JobTarget(runner=runner, flavor="a10g-small") if runner == "hf_cloud" else None + cmd = build_training_command(request, "/tmp/out", job_target=target) + monkeypatch.setattr(sys, "argv", ["lerobot-train", *cmd[3:]]) + + cfg = _parse_training() + + assert cfg.policy.type == policy_type + assert cfg.trackio.space_id == "me/lelab-trackio" + + +def test_rollout_argv_parses(monkeypatch, tmp_path) -> None: + import lerobot.scripts.lerobot_rollout as lerobot_rollout + from lelab import rollout + from lelab.rollout import InferenceRequest + from lerobot.configs.types import FeatureType, PolicyFeature + from lerobot.policies.act.configuration_act import ACTConfig + + checkpoint = tmp_path / "checkpoints" / "000050" / "pretrained_model" + checkpoint.mkdir(parents=True) + ACTConfig( + input_features={ + "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(6,)), + "observation.images.front": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 480, 640)), + }, + output_features={"action": PolicyFeature(type=FeatureType.ACTION, shape=(6,))}, + )._save_pretrained(checkpoint) + + commands: list[list[str]] = [] + + def capture(cmd, **kwargs): + commands.append(cmd) + raise OSError("stop before spawning") + + monkeypatch.setattr(rollout, "inference_active", False) + monkeypatch.setattr(rollout, "setup_follower_calibration_file", lambda name: "robot_a") + monkeypatch.setattr(rollout.subprocess, "Popen", capture) + rollout.handle_start_inference( + InferenceRequest( + follower_port="/dev/ttyUSB0", + follower_config="robot_a", + policy_ref=str(checkpoint), + task="step 1: grab the cube", + cameras={"front": {"type": "opencv", "camera_index": 0, "width": 640, "height": 480, "fps": 30}}, + ) + ) + + parsed = [] + + class _StopBeforeHardwareError(Exception): + pass + + def stop_before_hardware(cfg, _shutdown_event): + parsed.append(cfg) + raise _StopBeforeHardwareError + + monkeypatch.setattr(lerobot_rollout, "build_rollout_context", stop_before_hardware) + monkeypatch.setattr(sys, "argv", ["lerobot-rollout", *commands[0][3:]]) + with pytest.raises(_StopBeforeHardwareError): + lerobot_rollout.main() + + (cfg,) = parsed + assert cfg.robot.port == "/dev/ttyUSB0" + assert cfg.robot.id == "robot_a" + assert cfg.task == "step 1: grab the cube" + assert set(cfg.robot.cameras) == {"front"} From 954d569da50fde8e9bfe1dc92ac29d3af5ac2034 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:22:23 +0100 Subject: [PATCH 11/12] refactor: log tracebacks through logger.exception LeRobot's init_logging() now installs a real logging.Formatter that keeps exc_info (huggingface/lerobot#4215), so the tracebacks folded into the message by hand are no longer needed. --- lelab/calibrate.py | 4 +--- lelab/record.py | 9 ++------- 2 files changed, 3 insertions(+), 10 deletions(-) diff --git a/lelab/calibrate.py b/lelab/calibrate.py index dd9fcefbb..a3f40620c 100644 --- a/lelab/calibrate.py +++ b/lelab/calibrate.py @@ -22,7 +22,6 @@ import logging import threading import time -import traceback from copy import deepcopy from dataclasses import dataclass from typing import Any, Literal @@ -313,8 +312,7 @@ def _calibration_worker(self, request: CalibrationRequest): self._update_status(error=str(e)) self._cleanup_and_finish(str(e), status="error") except Exception as e: - logger.error(f"Calibration error: {e}") - logger.error(traceback.format_exc()) + logger.exception(f"Calibration error: {e}") # Ensure cleanup happens even on error self._cleanup_and_finish(f"Calibration failed: {e}", status="error") finally: diff --git a/lelab/record.py b/lelab/record.py index af731cbb0..f48724e2e 100644 --- a/lelab/record.py +++ b/lelab/record.py @@ -17,7 +17,6 @@ import shutil import threading import time -import traceback from datetime import datetime from typing import Any @@ -347,10 +346,7 @@ def recording_worker(): "robot_type": getattr(dataset.meta, "robot_type", "Unknown robot"), } except Exception as e: - # lerobot's init_logging() installs a root formatter that only - # renders the message, so logger.exception() would drop the - # traceback. Fold it into the message instead. - logger.error("Recording session failed: %s\n%s", e, traceback.format_exc()) + logger.exception("Recording session failed: %s", e) current_phase = "error" if recording_start_time: session_end_elapsed_seconds = int(time.time() - recording_start_time) @@ -760,8 +756,7 @@ def handle_upload_dataset(request: UploadRequest) -> dict[str, Any]: return {"success": False, "message": str(e)} except Exception as e: - logger.error(f"Error uploading dataset {request.dataset_repo_id}: {e}") - logger.error(f"Full traceback: {traceback.format_exc()}") + logger.exception(f"Error uploading dataset {request.dataset_repo_id}: {e}") err_text = str(e).lower() looks_like_auth = any( From f69ba57dbd1a50daf90c5b176fce5185c5d67d15 Mon Sep 17 00:00:00 2001 From: Nicolas Rabault Date: Thu, 5 Nov 2026 16:23:01 +0100 Subject: [PATCH 12/12] docs: describe the LeRobot v0.7.0 pin, motor bus and CLI test in CLAUDE.md Also fix two module descriptions that no longer matched the code: recording drives record_loop with its own events dict rather than patching LeRobot's keyboard listener, and teleoperation runs its own loop. --- CLAUDE.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/CLAUDE.md b/CLAUDE.md index 079894fbe..58c5e2dd7 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -16,7 +16,7 @@ Install (editable, requires Python ≥3.10): pip install -e . ``` -`lerobot` is pinned to a release tag on `huggingface/lerobot` (currently `v0.6.1`, see [pyproject.toml](pyproject.toml)) — installed from git rather than PyPI because the published wheels don't expose the `lerobot-rollout` script that [lelab/rollout.py](lelab/rollout.py) shells out to. Bump the tag when you want a newer upstream release; expect import-path drift (e.g. `lerobot.configs.dataset`, `lerobot.utils.feature_utils`, `lerobot.common.control_utils`) and adjust call sites accordingly. +`lerobot` is pinned to a ref on `huggingface/lerobot` (currently `v0.7.0`, see [pyproject.toml](pyproject.toml)), installed from git rather than PyPI because the published wheels don't expose the `lerobot-rollout` script that [lelab/rollout.py](lelab/rollout.py) shells out to. Bump the ref when you want a newer upstream release; expect import-path drift (e.g. `lerobot.configs.dataset`, `lerobot.utils.feature_utils`, `lerobot.common.control_utils`) and CLI flag renames (v0.7.0 replaced `--wandb.*` with `--trackio.*`). [tests/test_lerobot_cli.py](tests/test_lerobot_cli.py) runs the training and rollout argv LeLab builds through the installed LeRobot's parsers, so run the tests right after a bump. Since v0.7.0 the arms' motors go through LeRobot's rustypot-backed `SerialMotorsBus`: use its public methods (`set_operating_mode`, `set_half_turn_homings`, `disconnect`), not Feetech SDK internals. Run servers (entry point defined in [pyproject.toml](pyproject.toml)): @@ -35,8 +35,8 @@ Run the Python tests with `pytest` (config in [pyproject.toml](pyproject.toml); [server.py](lelab/server.py) is a thin FastAPI router. Each feature lives in its own module that owns its global state and exposes `handle_*` functions plus a Pydantic request model: -- [record.py](lelab/record.py) — dataset recording (wraps `lerobot.record.record`); patches `lerobot.common.utils.control_utils` keyboard listener so frontend buttons replace arrow-key controls. The worker is a daemon thread, so the server's shutdown hook calls `stop_recording_and_wait()` to let an active session finalize on Ctrl+C. -- [teleoperate.py](lelab/teleoperate.py) — leader→follower teleoperation (wraps `lerobot.teleoperate`). +- [record.py](lelab/record.py): dataset recording around LeRobot's `record_loop`; frontend buttons set the same `events` dict LeRobot's keyboard listener would. The worker is a daemon thread, so the server's shutdown hook calls `stop_recording_and_wait()` to let an active session finalize on Ctrl+C. +- [teleoperate.py](lelab/teleoperate.py): leader→follower teleoperation loop over LeRobot's SO-101 leader and follower. - [replay.py](lelab/replay.py): plays a recorded episode's actions on the follower arm, an in-process take on `lerobot-replay` so the dataset browser can stop it. - [calibrate.py](lelab/calibrate.py) — step-by-step web calibration with a `CalibrationManager` singleton and `_step_complete` threading.Event. - [train.py](lelab/train.py) — wraps the LeRobot training CLI as a subprocess (psutil for lifecycle, queue for log streaming).