Add configuration files for ACT inference and training，add AAO simulation environment with lerobot-eval,implement rot6d convention wrappers and update AAOSimEnv for action/observation rotation handling

airbot_ie/lerobot_plugin/configs/act_infer_aao.yaml

@@ -0,0 +1,30 @@
+# ACT inference on AAO MuJoCo simulator via lerobot-rollout + aao_sim robot plugin.
+#
+# This uses the absolute_random_wrist checkpoint trained with observation_rotation=rot6d
+# (10-dim state = pos[3] + rot6d[6] + gripper[1]). The env must match by setting
+# observation_rotation: rot6d below.
+#
+# Run:
+#   pixi run -e infer-aao mcap-lerobot-infer -c airbot_ie/lerobot_plugin/configs/act_infer_aao.yaml
+
+preserve_keys: [sim_cameras]
+
+robot:
+  type: aao_sim
+  task_config: open_door
+  operator: arm
+  kinematic: false
+  substeps: 1
+  has_gripper: true
+  observation_rotation: rot6d      # → 10-dim state = pos(3) + rot6d(6) + grip(1)
+  sim_cameras:
+    wrist_cam: wrist_cam
+  camera_shape: [352, 640, 3]
+  mujoco_gl: egl
+
+policy:
+  path: models_open_door/absolute_random_wrist/pretrained_model
+
+task: "open the door"
+fps: 30
+duration: 10

airbot_ie/lerobot_plugin/configs/act_infer_mock.yaml

@@ -0,0 +1,36 @@
+# ACT inference on MOCK AIRBOT Play arm (no hardware needed).
+#
+# Uses pose control mode with arm + eef components:
+#   observation.state = eef pos(3) + quat(4) + gripper(1) = (8,)
+#   action            = target eef pos(3) + quat(4) + gripper(1) = (8,)
+#
+# Run:
+#   pixi run -e act-infer mcap-lerobot-infer -c airbot_ie/lerobot_plugin/configs/act_infer_mock.yaml
+
+preserve_keys: [cameras]
+
+robot:
+  type: airbot_play
+  mock: true
+  control_mode: pose
+  components: [arm, eef]       # observation.state (8,): eef pos+quat+gripper
+  cameras:
+    eef_wrist_cam:
+      target: airdc.common.devices.cameras.mock.MockCamera
+      width: 640
+      height: 352
+      fps: 30
+      extra: { random: true }
+    env1_cam:
+      target: airdc.common.devices.cameras.mock.MockCamera
+      width: 640
+      height: 352
+      fps: 30
+      extra: { random: true }
+
+policy:
+  path: outputs/train/my_act/checkpoints/last/pretrained_model
+
+task: "open the door"
+fps: 30
+duration: 3

airbot_ie/lerobot_plugin/configs/act_infer_real.yaml

@@ -0,0 +1,39 @@
+# ACT inference on REAL AIRBOT Play arm.
+#
+# Before running: update `url` and camera `camera_index` values for your setup.
+# Uses pose control mode with arm + eef components:
+#   observation.state = eef pos(3) + quat(4) + gripper(1) = (8,)
+#   action            = target eef pos(3) + quat(4) + gripper(1) = (8,)
+#
+# Run:
+#   pixi run -e act-infer mcap-lerobot-infer -c airbot_ie/lerobot_plugin/configs/act_infer_real.yaml
+
+preserve_keys: [cameras]
+
+robot:
+  type: airbot_play
+  control_mode: pose
+  components: [arm, eef]       # observation.state (8,): eef pos+quat+gripper
+  url: localhost               # ← change to arm's IP if remote
+  port: 50051
+  return_to_initial_position: true
+  cameras:
+    eef_wrist_cam:
+      target: airdc.common.devices.cameras.v4l2.V4L2Camera
+      camera_index: 0          # ← change to your wrist camera index
+      width: 640
+      height: 352
+      fps: 30
+    env1_cam:
+      target: airdc.common.devices.cameras.v4l2.V4L2Camera
+      camera_index: 1          # ← change to your env camera index
+      width: 640
+      height: 352
+      fps: 30
+
+policy:
+  path: outputs/train/my_act/checkpoints/last/pretrained_model
+
+task: "open the door"
+fps: 30
+duration: 0                    # 0 = run until Ctrl-C

airbot_ie/lerobot_plugin/configs/act_train.yaml

@@ -0,0 +1,57 @@
+# ACT (Action Chunking with Transformers) training config on MCAP dataset.
+#
+# ACT is a lightweight policy (no pretrained backbone), trained from scratch on
+# your own demonstrations. Much smaller than pi0.5/SmolVLA and faster to train.
+#
+# Run:
+#   pixi run -e act mcap-lerobot-train -c airbot_ie/lerobot_plugin/configs/act_train.yaml
+
+env:
+  HF_HUB_OFFLINE: "1"
+  TRANSFORMERS_OFFLINE: "1"
+
+# ============================================================================
+# Common training knobs
+# ============================================================================
+steps: 20000
+save_freq: 5000
+log_freq: 100
+batch_size: 8
+num_workers: 1
+output_dir: outputs/train/my_act
+seed: 1000
+
+wandb:
+  enable: true
+  project: airbot-act
+  mode: offline
+
+policy:
+  type: act
+  push_to_hub: false
+  chunk_size: 100          # ACT action horizon
+  n_action_steps: 100      # steps to execute per chunk
+
+dataset:
+  root: data
+  repo_id: mcap
+  streaming: true
+
+mcap:
+  # observation.state = arm eef pose (position[3] + orientation[4]) + gripper (1) -> (8,)
+  states:
+    - arm/pose/position
+    - arm/pose/orientation
+    - eef/joint_state/position
+  # Wrist cam + env cam -> each (3, H, W)
+  images:
+    - eef_wrist_cam/color/image_raw
+    - env1_cam/color/image_raw
+  # action = target eef pose (position[3] + orientation[4]) + gripper (1) -> (8,)
+  actions:
+    - action/arm/pose/position
+    - action/arm/pose/orientation
+    - action/eef/joint_state/position
+  # ACT does not use language conditioning
+  task_source: none
+  task: "do the task"

airbot_ie/lerobot_plugin/lerobot_env_aao_sim/README.md

@@ -0,0 +1,65 @@
+# lerobot_env_aao_sim
+
+LeRobot **Env** plugin: a `gymnasium.Env` over the auto-atomic-operation (aao)
+MuJoCo simulator, for **native success-rate evaluation** with `lerobot-eval`.
+
+## Why this exists
+
+`lerobot-rollout` (the Robot path used by `lerobot_robot_aao_sim`) is written for
+real hardware — it has no success detection and just runs to `duration`.
+`lerobot-eval` drives a gymnasium `VectorEnv` and computes `pc_success` natively
+from each env's `step()` returning `info["is_success"]`. This plugin supplies
+that env, so you get success rate + mp4 videos out of the box.
+
+LeRobot auto-discovers this by the distribution name prefix `lerobot_env_`, and
+`@EnvConfig.register_subclass("aao_sim")` registers `--env.type=aao_sim`.
+
+## Success signal
+
+`step()` reports success via `backend.is_object_displaced(success_object,
+initial_pose, threshold)` — the same displacement check the aao task graph uses.
+For `open_door` the target is `handle_body_phys` (threshold 0.01 m). Both are
+config fields (`success_object`, `displacement_threshold`).
+
+## Usage
+
+```bash
+pixi run -e infer-aao lerobot-eval \
+  --policy.path=models_open_door/absolute_random_wrist/pretrained_model \
+  --env.type=aao_sim \
+  --env.task_config=open_door \
+  --env.observation_rotation=rot6d \
+  --eval.n_episodes=10 \
+  --eval.batch_size=1 \
+  --policy.device=cuda
+```
+
+Outputs: `outputs/eval/<run>/videos/*.mp4` + `metrics.json` with `pc_success`.
+
+## observation_rotation
+
+| Value | observation.state dim | Use when |
+|---|---|---|
+| `rot6d` (default) | 10 (pos3 + rot6d6 + grip1) | checkpoint trained with 6-D obs (e.g. absolute_random_wrist) |
+| `quat` | 8 (pos3 + quat4 + grip1) | checkpoint trained with quaternion obs |
+
+`action_rotation` controls the policy action representation:
+
+| Value | action dim | Execution path |
+|---|---|---|
+| `quat` (default) | 8 (pos3 + quat4 + grip1) | passed directly to `apply_pose_action` |
+| `rot6d` | 10 (pos3 + rot6d6 + grip1) | converted to quaternion before `apply_pose_action` |
+
+## rot6d convention
+
+AAO simulation, MCAP `Rotation6D.quat_to_rot6d()` conversion, and real-robot
+data all use the same column-vector convention:
+
+```text
+[r00, r10, r20, r01, r11, r21]
+```
+
+It is the first two columns of the 3×3 rotation matrix. No observation or action
+convention wrapper is required. Checkpoints trained with the old row-major
+simulation representation are not compatible with this unified path and must
+be converted or retrained.

airbot_ie/lerobot_plugin/lerobot_env_aao_sim/lerobot_env_aao_sim/__init__.py

@@ -0,0 +1,14 @@
+"""LeRobot Env plugin: a gymnasium env over the auto-atomic-operation MuJoCo
+simulator, for native success-rate evaluation via ``lerobot-eval``.
+
+Unlike ``lerobot_robot_aao_sim`` (which drives ``lerobot-rollout`` and has no
+success signal), this exposes a ``gym.Env`` whose ``step()`` returns
+``info["is_success"]`` so ``lerobot-eval`` computes ``pc_success`` natively.
+
+LeRobot auto-discovers this by the distribution name prefix ``lerobot_env_``,
+and ``@EnvConfig.register_subclass("aao_sim")`` registers ``--env.type=aao_sim``.
+"""
+
+from .config_aao_sim_env import AAOSimEnvConfig
+
+__all__ = ["AAOSimEnvConfig"]