Repository navigation
Conversation
…NI sweep
Immediate action (docs):
- Fix the crawl-slope WBT example that recommended sampling_mode=start
(unsupported by the tensor sampler; raises NotImplementedError at env
construction): use sampling_mode=mixed with sampling_start_ratio=1.0, the
supported deterministic clip-start form, in both en and zh_CN pages.
- Narrow MotionCommandCfg.sampling_mode to Literal["adaptive", "mixed"] so
invalid values fail at config validation.
High-confidence deletions, batch 1 (zero in-repo consumers):
- managers: clear_finite_verdict x3, TerminationManager.dones/get_term,
RecorderManager.__contains__ (and Null), DelayBuffer.set_lags,
apply_cuda_process_sharing_manifest, get_hydra_runtime_choice.
- base: EnvConfigAdapter algo_name (silently ignored; all call sites cleaned,
including build_play_backend_adapter), registry.find_available_sim_backend
(make() already inlines the fallback).
- envs/tasks: UniformVelocityCommandCfg VizCfg/viz, BaseActionCfg.preserve_order,
MotionLoader.get_clip_indices, resolve_checkpoint_file Sequence[str] branch.
- rl: resolve_torch_device_alias group in rl.runtime.device, build_actor
num_envs (silently ignored).
- conf: PPO algo.experiment_name/run_name, SAC algo.actor: {} fake-compat key,
motion-tracking owner YAML lines that only restate dataclass defaults.
High-confidence deletions, batch 2 (issue comments: downstream audit upgraded
all maintainer-decision items):
- UniformVelocityCommandCfg.init_velocity_prob fake knob (8 owner YAML lines
removed; downstream repos must drop their init_velocity_prob: 0.0 lines on
upgrade, tracked in the issue).
- EnvAlgoCapabilities action_low/action_high/joint_names reserved fields.
- Tensor-migration leftovers: motion_tracking/common/tensor_runtime.py,
tensor_state_store.py (DeviceStateStore), locomotion/go2/tensor_parity.py;
ADR-0011/0012 evidence pointers updated.
- PPO owner algo.resume/algo.resume_path dead keys; the PPO entrypoint fails
closed via reject_removed_ppo_resume_keys (resume stays on
algo.load_run + algo.checkpoint; off-policy owners keep the live keys).
- TorchEnv.record_finite_validation label argument (silently dropped).
- Tests-only debug surface: ManagerBase.get_active_iterable_terms and
implementations, extract_reset_obs, render_play_mode and
run_native_viewer_play_loop thin wrappers (callers use env.run_playback
directly), OnPolicyLogger (with its api_reference entry).
TorchEnv.require is kept: it is the live deferred mask-check sink installed
by ManagerBasedRlEnv for masked autoreset.
- Zero-consumer noise cfgs ConstantNoiseCfg/GaussianNoiseCfg/
SegmentwiseUniformNoiseCfg (UniformNoiseCfg and the NoiseModel extension
base stay).
Validation: make check; make test (2946 passed, 67 skipped).
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概要
执行 issue #2629 的「立即行动」与全部「高置信可删」条目(含 issue 评论中经下游生态审计复核后由"维护者决策"升级的批二条目)。97 个文件,+258/−3429。
立即行动:文档错误推荐
2-motion_tracking.md(en + zh_CN)的 crawl-slope WBT 示例原推荐sampling_mode=start,按文档操作会在 env 构造期抛NotImplementedError(tensor sampler 从未支持该值)。改为受支持的确定性 clip 起点形式sampling_mode=mixed+sampling_start_ratio=1.0。MotionCommandCfg.sampling_mode收窄为Literal["adaptive", "mixed"],非法值在配置校验期即被清晰拒绝。批一:高置信可删(issue 正文,18-20 处)
clear_finite_verdict×3、TerminationManager.dones/get_term、RecorderManager.__contains__(含 Null)、DelayBuffer.set_lags、apply_cuda_process_sharing_manifest、get_hydra_runtime_choice(及unilab.trainingre-export)。EnvConfigAdapter的algo_name(静默忽略;~20 处调用点全部清理,含build_play_backend_adapter转发参数)、registry.find_available_sim_backend(make()已内联同一回退)。UniformVelocityCommandCfg.VizCfg/viz、BaseActionCfg.preserve_order、MotionLoader.get_clip_indices、resolve_checkpoint_file的Sequence[str]分支。rl.runtime.device的resolve_torch_device_alias整组(含 4 个私有助手)、build_actor的num_envs。algo.experiment_name/algo.run_name、SACalgo.actor: {}假兼容键、4 个 motion-tracking owner YAML 中仅重复 dataclass 默认值的 3 行。批二:评论升级项(原"维护者决策"全部转删除)
init_velocity_prob假旋钮:字段、fail-closed 分支、8 个 owner YAML 行删除。跨仓协调:5 个下游仓库(unitree/engineai 等)的 10 行init_velocity_prob: 0.0需在升级时同步删除(issue 已跟踪)。EnvAlgoCapabilities的action_low/action_high/joint_names预留字段(producer 恒填 None、零消费)。tensor_runtime.py、tensor_state_store.py::DeviceStateStore、go2/tensor_parity.py及其专属测试;ADR-0011/0012 的 Evidence 引用已同步更新。algo.resume/algo.resume_path:读点仅在 off-policy/APPO 路径。删除并新增reject_removed_ppo_resume_keys(PPO 入口 fail-closed;PPO resume 仍走algo.load_run+algo.checkpoint;off-policy owner 保留活键)。TorchEnv.record_finite_validation的label参数(实现体del label静默丢弃):收窄签名,5 个生产调用点与协议定义同步更新。ManagerBase.get_active_iterable_terms及 9 个实现、extract_reset_obs、render_play_mode/run_native_viewer_play_loop薄封装(调用方与scripts/visualize_task_env.py改为直接使用env.run_playback,测试契约覆盖保留)、OnPolicyLogger(含 api_reference 条目)。TorchEnv.require原计划删除,但实际是 masked autoreset 延迟 mask 校验的活 sink(set_deferred_mask_checks(self.require)),已保留并恢复其测试。ConstantNoiseCfg/GaussianNoiseCfg/SegmentwiseUniformNoiseCfg删除(生产 YAML 只用UniformNoiseCfg);UniformNoiseCfg与NoiseModel/NoiseModelWithAdditiveBiasCfg扩展基座保留。验证
make check✅(ruff format/check、mypy、tests 静态检查)make test✅ 2946 passed, 67 skipped, 0 failedrg全仓(src/tests/scripts/docs)复核零消费;下游 5 仓审计结论来自 issue 评论。备注
docs/sphinx/source/changelog.md的 Unreleased(Fixed + Removed)添加双语条目。make test-all(含 coverage/benchmark smoke);如需要可补跑。