get-embodied-tasks 0.1.0__tar.gz

get_embodied_tasks-0.1.0/.gitignore

@@ -0,0 +1,8 @@
+__pycache__/
+*.pyc
+*.egg-info/
+.pytest_cache/
+.ruff_cache/
+build/
+dist/
+.venv/

get_embodied_tasks-0.1.0/PKG-INFO

@@ -0,0 +1,88 @@
+Metadata-Version: 2.4
+Name: get-embodied-tasks
+Version: 0.1.0
+Summary: Resolve an embodied-AI task description to its dataset and ready-to-run train/eval commands across pluggable backends (LeRobot, ...).
+Project-URL: Homepage, https://github.com/OpenGHz/get-embodied-tasks
+Project-URL: Repository, https://github.com/OpenGHz/get-embodied-tasks
+Author: OpenGHz
+License: Apache-2.0
+Keywords: agent-skill,benchmark,datasets,embodied-ai,lerobot,robotics
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+# get-embodied-tasks
+
+Resolve an embodied-AI task description to its **existing dataset** and
+**ready-to-run train/eval commands**, across pluggable backends. **LeRobot** is the
+first backend; more can be added. Ships as both a CLI and an agent skill (Claude +
+Codex).
+
+```bash
+get-embodied-tasks "stack blocks"
+#  → [lerobot] RoboTwin stack_blocks_two
+#    Dataset: lerobot/robotwin_unified
+#    Train:   lerobot-train --policy.type=smolvla --dataset.repo_id=lerobot/robotwin_unified ...
+#    Eval:    lerobot-eval  --env.type=robotwin --env.task=stack_blocks_two ...
+```
+
+## Design
+
+- **Facts in the tool, judgment in the skill.** The CLI emits real dataset ids and
+  command flags (never model-recalled). The `SKILL.md` playbook parses a fuzzy task
+  description, calls the CLI, and on a miss researches alternatives or drafts a plan.
+- **Backends are read, not vendored.** The LeRobot adapter reads a *vanilla* LeRobot
+  checkout by AST over its `envs/` source (no import of torch/gymnasium; no patches to
+  LeRobot). RoboTwin/RoboMME task descriptions ship here in `data/`.
+- **Drift is surfaced, never hidden.** If a backend's expected constant/file changes,
+  the catalog emits structured warnings (stderr, `--json` `warnings`, `--strict`,
+  `doctor`) so it can be fixed.
+
+## Install
+
+```bash
+pip install -e .
+bash install.sh        # symlink into ~/.agents/skills (Claude + Codex pick it up)
+```
+
+## Backend discovery
+
+For each backend, in order: `$<ENV_VAR>` → `./third_party/ei_ws/<name>` →
+`~/ei_ws/<name>` → installed package. For LeRobot set `LEROBOT_HOME=<checkout>` or
+just `pip install lerobot`. Configure backends in
+[`src/get_embodied_tasks/backends.toml`](src/get_embodied_tasks/backends.toml).
+
+## CLI
+
+```
+get-embodied-tasks "<query>" [-b BENCHMARK] [--backend NAME] [-n N] [--policy P] [--json]
+get-embodied-tasks --benchmark <b>     # list a benchmark
+get-embodied-tasks --dump              # whole catalog as JSON
+get-embodied-tasks doctor              # backend health / drift report
+```
+
+## Reference docs (per backend)
+
+[`docs/backends/lerobot/`](docs/backends/lerobot/) holds reference material for the
+LeRobot backend, regeneratable from a vanilla checkout:
+
+- [`cli_reference.md`](docs/backends/lerobot/cli_reference.md) — every `lerobot-*` command + its extra (generated)
+- [`components_reference.md`](docs/backends/lerobot/components_reference.md) — registered policies/robots/teleops/cameras/optimizers/schedulers/processors (generated)
+- [`troubleshooting.md`](docs/backends/lerobot/troubleshooting.md) — consolidated FAQ (hand-written, links upstream)
+
+Refresh the generated ones:
+
+```bash
+python docs/scripts/gen_lerobot_reference.py --lerobot-root <checkout>   # or rely on discovery
+python docs/scripts/gen_lerobot_reference.py --check                     # CI: fail if stale
+```
+
+## Adding a backend
+
+Add an adapter under `src/get_embodied_tasks/backends/` implementing
+`backends/base.py:Backend`, register it in `cli.py:_ADAPTERS`, and add a table to
+`backends.toml`.

get_embodied_tasks-0.1.0/README.md

@@ -0,0 +1,71 @@
+# get-embodied-tasks
+
+Resolve an embodied-AI task description to its **existing dataset** and
+**ready-to-run train/eval commands**, across pluggable backends. **LeRobot** is the
+first backend; more can be added. Ships as both a CLI and an agent skill (Claude +
+Codex).
+
+```bash
+get-embodied-tasks "stack blocks"
+#  → [lerobot] RoboTwin stack_blocks_two
+#    Dataset: lerobot/robotwin_unified
+#    Train:   lerobot-train --policy.type=smolvla --dataset.repo_id=lerobot/robotwin_unified ...
+#    Eval:    lerobot-eval  --env.type=robotwin --env.task=stack_blocks_two ...
+```
+
+## Design
+
+- **Facts in the tool, judgment in the skill.** The CLI emits real dataset ids and
+  command flags (never model-recalled). The `SKILL.md` playbook parses a fuzzy task
+  description, calls the CLI, and on a miss researches alternatives or drafts a plan.
+- **Backends are read, not vendored.** The LeRobot adapter reads a *vanilla* LeRobot
+  checkout by AST over its `envs/` source (no import of torch/gymnasium; no patches to
+  LeRobot). RoboTwin/RoboMME task descriptions ship here in `data/`.
+- **Drift is surfaced, never hidden.** If a backend's expected constant/file changes,
+  the catalog emits structured warnings (stderr, `--json` `warnings`, `--strict`,
+  `doctor`) so it can be fixed.
+
+## Install
+
+```bash
+pip install -e .
+bash install.sh        # symlink into ~/.agents/skills (Claude + Codex pick it up)
+```
+
+## Backend discovery
+
+For each backend, in order: `$<ENV_VAR>` → `./third_party/ei_ws/<name>` →
+`~/ei_ws/<name>` → installed package. For LeRobot set `LEROBOT_HOME=<checkout>` or
+just `pip install lerobot`. Configure backends in
+[`src/get_embodied_tasks/backends.toml`](src/get_embodied_tasks/backends.toml).
+
+## CLI
+
+```
+get-embodied-tasks "<query>" [-b BENCHMARK] [--backend NAME] [-n N] [--policy P] [--json]
+get-embodied-tasks --benchmark <b>     # list a benchmark
+get-embodied-tasks --dump              # whole catalog as JSON
+get-embodied-tasks doctor              # backend health / drift report
+```
+
+## Reference docs (per backend)
+
+[`docs/backends/lerobot/`](docs/backends/lerobot/) holds reference material for the
+LeRobot backend, regeneratable from a vanilla checkout:
+
+- [`cli_reference.md`](docs/backends/lerobot/cli_reference.md) — every `lerobot-*` command + its extra (generated)
+- [`components_reference.md`](docs/backends/lerobot/components_reference.md) — registered policies/robots/teleops/cameras/optimizers/schedulers/processors (generated)
+- [`troubleshooting.md`](docs/backends/lerobot/troubleshooting.md) — consolidated FAQ (hand-written, links upstream)
+
+Refresh the generated ones:
+
+```bash
+python docs/scripts/gen_lerobot_reference.py --lerobot-root <checkout>   # or rely on discovery
+python docs/scripts/gen_lerobot_reference.py --check                     # CI: fail if stale
+```
+
+## Adding a backend
+
+Add an adapter under `src/get_embodied_tasks/backends/` implementing
+`backends/base.py:Backend`, register it in `cli.py:_ADAPTERS`, and add a table to
+`backends.toml`.

get_embodied_tasks-0.1.0/SKILL.md

@@ -0,0 +1,105 @@
+---
+name: get-embodied-tasks
+description: >-
+  For an embodied-AI / robot-manipulation task description (scene, device, objects,
+  operation, steps, perturbation, memory, etc.), find the matching existing simulation
+  task across pluggable backends (LeRobot today, more later) — its dataset (download
+  link) and the exact train/eval commands to run it. If nothing fits, research other
+  open-source projects/datasets, and if none exist, draft an implementation plan. Use
+  when the user asks "which dataset/benchmark has task X", "how do I train/eval on X",
+  "给我这个任务的数据集和训练/推理命令", "哪个 benchmark 有这个任务",
+  "找一个搭积木/带光照扰动/需要记忆的数据集", or describes a manipulation task and wants
+  data + runnable configs.
+---
+
+# get-embodied-tasks
+
+Turn a task description into **(1) the existing dataset + download link** and
+**(2) ready-to-run training/evaluation commands**, sourced from real backends. When
+no backend covers it, **research** alternatives or **draft an implementation plan**.
+
+## Core principle
+
+**Facts come from the tool, judgment comes from you.** Dataset repo ids and command
+flags MUST come verbatim from the `get-embodied-tasks` CLI — never invent a dataset
+id or a command from memory. Your job: translate the description into the right
+query/benchmark, pick among results, and explain.
+
+Scope: **simulation** tasks that already have datasets. Backends are discovered, not
+hard-coded (see "Backends" below).
+
+## Workflow
+
+### 1. Parse the description → query + facet
+
+Pull the salient noun/verb ("stack blocks", "insert peg", "open drawer"). Map
+distinctive requirements to a benchmark with `--benchmark`:
+
+| Mentions… | `--benchmark` |
+| --- | --- |
+| memory / counting / occlusion / history-dependent | `robomme` |
+| perturbation / robustness / viewpoint / lighting / distractor | `libero_plus` |
+| bimanual / dual-arm / handover | `robotwin` |
+| language-conditioned multi-task tabletop | `libero`, `metaworld` |
+| kitchen / household | `robocasa` |
+| long-horizon reasoning, tool use | `vlabench` |
+
+If unsure, omit `--benchmark` (searches everything; a task may match several
+benchmarks — present the best fit and note the alternatives).
+
+### 2. Call the tool (source of truth)
+
+```bash
+get-embodied-tasks "<query>"                 # dataset + train/eval per match
+get-embodied-tasks "<query>" --benchmark <b> # scope to a benchmark
+get-embodied-tasks --benchmark <b>           # list a whole benchmark
+get-embodied-tasks "<query>" --json          # structured
+get-embodied-tasks --dump                    # whole catalog as JSON (reason over it yourself)
+get-embodied-tasks doctor                    # backend health / drift check
+```
+
+Install if missing: `pip install -e /home/ghz/Work/skills/get-embodied-tasks`. It is
+stdlib-only (no torch/GPU) and reads the backend in place.
+
+### 3. Present the result
+
+For each chosen task: the task + description, the **dataset link**, and the **train**
+and **eval** commands exactly as emitted. Note the policy is a swappable default
+(`--policy.type=`) and that `lerobot-eval` is the sim "inference".
+
+### 4. Handle diagnostics (drift) — don't ignore warnings
+
+If the CLI prints `⚠ Diagnostics` (stderr) or `--json` has a non-empty `warnings`
+array, a backend's catalog drifted (e.g. LeRobot renamed/moved a constant, or the
+dataset map is stale). **Tell the user what drifted** (the warning names the file/
+constant) and **offer to fix it**: the fix is an edit to the backend adapter
+(`src/get_embodied_tasks/catalog.py` constant names, or `data/lerobot_datasets.py`).
+Run `get-embodied-tasks doctor` for the full health report. Do not silently proceed
+on partial data.
+
+For reference while fixing (current registry/CLI names of the LeRobot backend), see
+`docs/backends/lerobot/{components_reference,cli_reference,troubleshooting}.md`;
+regenerate the first two with `python docs/scripts/gen_lerobot_reference.py`.
+
+### 5. Fallback — when nothing fits
+
+1. **Research** other open-source benchmarks/datasets (use the `deep-research` skill
+   or web search). Report concrete links, not guesses.
+2. **If none exist**, draft an **implementation plan** (delegate to the `Plan` agent):
+   how to get data (generate in sim / record), which policy, the train/eval setup.
+   State plainly that the task is not currently covered.
+
+## Backends
+
+`get-embodied-tasks` is backend-pluggable; **LeRobot** is the first backend. Each
+backend is located by discovery: `$<ENV_VAR>` → `./third_party/ei_ws/<name>` →
+`~/ei_ws/<name>` → installed package. For LeRobot, point `LEROBOT_HOME` at a checkout
+or just have `lerobot` installed. Adding a backend = a new adapter under
+`src/get_embodied_tasks/backends/` + an entry in `backends.toml`.
+
+## Don'ts
+
+- Don't output a dataset id or command the CLI didn't emit.
+- Don't claim a task is covered when there was no match — go to step 5.
+- Don't ignore `⚠ Diagnostics` — surface and offer to fix (step 4).
+- Sim-only: no real-robot `lerobot-record`/`lerobot-rollout` commands.

get_embodied_tasks-0.1.0/docs/backends/lerobot/cli_reference.md

@@ -0,0 +1,26 @@
+# LeRobot CLI Reference
+
+<!-- AUTO-GENERATED by docs/scripts/gen_lerobot_reference.py; do not edit by hand. Run it to refresh. -->
+
+All `lerobot-*` commands of the discovered LeRobot checkout. Run any with `--help`.
+Install a command's extra with `pip install 'lerobot[<extra>]'`.
+
+| Command | Installs with | Summary |
+|---|---|---|
+| `lerobot-calibrate` | `hardware` | Helper to recalibrate your device (robot or teleoperator). |
+| `lerobot-dataset-viz` | `dataset` | Visualize data of **all** frames of any episode of a dataset of type LeRobotDataset. |
+| `lerobot-edit-dataset` | `dataset` | Edit LeRobot datasets using various transformation tools. |
+| `lerobot-eval` | `evaluation` | Evaluate a policy on an environment by running rollouts and computing metrics. |
+| `lerobot-find-cameras` | `hardware` | Helper to find the camera devices available in your system. |
+| `lerobot-find-joint-limits` | `hardware` | Script to find joint limits and end-effector bounds via teleoperation. |
+| `lerobot-find-port` | `hardware` | Helper to find the USB port associated with your MotorsBus. |
+| `lerobot-imgtransform-viz` | `dataset` | Visualize effects of image transforms for a given configuration. |
+| `lerobot-info` | `—` | Use this script to get a quick summary of your system config. |
+| `lerobot-record` | `core_scripts` | Records a dataset via teleoperation.  This is a pure data-collection |
+| `lerobot-replay` | `core_scripts` | Replays the actions of an episode from a dataset on a robot. |
+| `lerobot-rollout` | `—` | Policy deployment engine with pluggable rollout strategies. |
+| `lerobot-setup-can` | `—` | Setup and debug CAN interfaces for Damiao motors (e.g., OpenArms). |
+| `lerobot-setup-motors` | `hardware` | Helper to set motor ids and baudrate. |
+| `lerobot-teleoperate` | `core_scripts` | Simple script to control a robot from teleoperation. |
+| `lerobot-train` | `training` | Train a policy. |
+| `lerobot-train-tokenizer` | `—` | Train FAST tokenizer for action encoding. |

get_embodied_tasks-0.1.0/docs/backends/lerobot/components_reference.md

@@ -0,0 +1,150 @@
+# LeRobot Components Reference
+
+<!-- AUTO-GENERATED by docs/scripts/gen_lerobot_reference.py; do not edit by hand. Run it to refresh. -->
+
+Every component registered in the LeRobot codebase, by kind. Generated by AST-scanning
+the registries of the discovered LeRobot checkout, so it matches that version.
+
+## Policies (16)
+
+Pass as `--policy.type=<type>` to `lerobot-train`, or load a trained one with `--policy.path=`.
+
+| Type | Class | Summary |
+|---|---|---|
+| `act` | `ACTConfig` | Configuration class for the Action Chunking Transformers policy. |
+| `diffusion` | `DiffusionConfig` | Configuration class for DiffusionPolicy. |
+| `eo1` | `EO1Config` | Configuration for native EO1 policy integration in LeRobot. |
+| `gaussian_actor` | `GaussianActorConfig` | Gaussian actor configuration. |
+| `groot` | `GrootConfig` | Configuration for Groot policy wrapper. |
+| `molmoact2` | `MolmoAct2Config` | MolmoAct2 policy backed by the converted HF checkpoint implementation. |
+| `multi_task_dit` | `MultiTaskDiTConfig` | Configuration for the Multi-Task Diffusion Transformer (DiT) policy. |
+| `pi0` | `PI0Config` |  |
+| `pi05` | `PI05Config` |  |
+| `pi0_fast` | `PI0FastConfig` |  |
+| `smolvla` | `SmolVLAConfig` |  |
+| `tdmpc` | `TDMPCConfig` | Configuration class for TDMPCPolicy. |
+| `vla_jepa` | `VLAJEPAConfig` |  |
+| `vqbet` | `VQBeTConfig` | Configuration class for VQ-BeT. |
+| `wall_x` | `WallXConfig` | Configuration class for Wall-X policy. |
+| `xvla` | `XVLAConfig` | Configuration class for the XVLA (Extended Vision-Language-Action) policy so it can |
+
+## Robots (17)
+
+Pass as `--robot.type=<type>` to `lerobot-record`, `lerobot-teleoperate`, `lerobot-calibrate`, etc.
+
+| Type | Class | Summary |
+|---|---|---|
+| `bi_openarm_follower` | `BiOpenArmFollowerConfig` | Configuration class for Bi OpenArm Follower robots. |
+| `bi_rebot_b601_follower` | `BiRebotB601FollowerConfig` | Configuration class for the bimanual reBot B601-DM follower robot. |
+| `bi_so_follower` | `BiSOFollowerConfig` | Configuration class for Bi SO Follower robots. |
+| `earthrover_mini_plus` | `EarthRoverMiniPlusConfig` | Configuration for EarthRover Mini Plus robot using Frodobots SDK. |
+| `hope_jr_arm` | `HopeJrArmConfig` |  |
+| `hope_jr_hand` | `HopeJrHandConfig` |  |
+| `koch_follower` | `KochFollowerConfig` |  |
+| `lekiwi` | `LeKiwiConfig` |  |
+| `lekiwi_client` | `LeKiwiClientConfig` |  |
+| `mujoco_env` | `MujocoEnvRobotConfig` | Configuration for wrapping a UnifiedMujocoEnv as a LeRobot robot. |
+| `omx_follower` | `OmxFollowerConfig` |  |
+| `openarm_follower` | `OpenArmFollowerConfig` |  |
+| `reachy2` | `Reachy2RobotConfig` |  |
+| `rebot_b601_follower` | `RebotB601FollowerRobotConfig` | Registered configuration for the reBot B601-DM follower robot. |
+| `so100_follower` | `SOFollowerRobotConfig` |  |
+| `so101_follower` | `SOFollowerRobotConfig` |  |
+| `unitree_g1` | `UnitreeG1Config` |  |
+
+## Teleoperators (19)
+
+Pass as `--teleop.type=<type>`.
+
+| Type | Class | Summary |
+|---|---|---|
+| `bi_openarm_leader` | `BiOpenArmLeaderConfig` | Configuration class for Bi OpenArm Follower robots. |
+| `bi_rebot_102_leader` | `BiRebotArm102LeaderConfig` | Configuration class for the bimanual reBot Arm 102 leader teleoperator. |
+| `bi_so_leader` | `BiSOLeaderConfig` | Configuration class for Bi SO Leader teleoperators. |
+| `gamepad` | `GamepadTeleopConfig` |  |
+| `homunculus_arm` | `HomunculusArmConfig` |  |
+| `homunculus_glove` | `HomunculusGloveConfig` |  |
+| `keyboard` | `KeyboardTeleopConfig` | KeyboardTeleopConfig |
+| `keyboard_ee` | `KeyboardEndEffectorTeleopConfig` | Configuration for keyboard end-effector teleoperator. |
+| `keyboard_rover` | `KeyboardRoverTeleopConfig` | Configuration for keyboard rover teleoperator. |
+| `koch_leader` | `KochLeaderConfig` |  |
+| `omx_leader` | `OmxLeaderConfig` |  |
+| `openarm_leader` | `OpenArmLeaderConfig` |  |
+| `openarm_mini` | `OpenArmMiniConfig` | Configuration for OpenArm Mini teleoperator with Feetech motors (dual arms). |
+| `phone` | `PhoneConfig` |  |
+| `reachy2_teleoperator` | `Reachy2TeleoperatorConfig` |  |
+| `rebot_102_leader` | `RebotArm102LeaderTeleopConfig` | Registered configuration for the reBot Arm 102 leader teleoperator. |
+| `so100_leader` | `SOLeaderTeleopConfig` |  |
+| `so101_leader` | `SOLeaderTeleopConfig` |  |
+| `unitree_g1` | `UnitreeG1TeleoperatorConfig` |  |
+
+## Cameras (4)
+
+Used inside `--robot.cameras="{ name: {type: <type>, ...} }"`.
+
+| Type | Class | Summary |
+|---|---|---|
+| `intelrealsense` | `RealSenseCameraConfig` | Configuration class for Intel RealSense cameras. |
+| `opencv` | `OpenCVCameraConfig` | Configuration class for OpenCV-based camera devices or video files. |
+| `reachy2_camera` | `Reachy2CameraConfig` | Configuration class for Reachy 2 camera devices. |
+| `zmq` | `ZMQCameraConfig` |  |
+
+## Optimizers (5)
+
+Pass as `--optimizer.type=<type>`.
+
+| Type | Class | Summary |
+|---|---|---|
+| `adam` | `AdamConfig` |  |
+| `adamw` | `AdamWConfig` |  |
+| `multi_adam` | `MultiAdamConfig` | Configuration for multiple Adam optimizers with different parameter groups. |
+| `sgd` | `SGDConfig` |  |
+| `xvla-adamw` | `XVLAAdamWConfig` | Custom AdamW optimizer for XVLA with differential learning rates. |
+
+## Schedulers (3)
+
+Pass as `--scheduler.type=<type>`.
+
+| Type | Class | Summary |
+|---|---|---|
+| `cosine_decay_with_warmup` | `CosineDecayWithWarmupSchedulerConfig` | Used by Physical Intelligence to train Pi0. |
+| `diffuser` | `DiffuserSchedulerConfig` |  |
+| `vqbet` | `VQBeTSchedulerConfig` |  |
+
+## Processor steps (31)
+
+Pipeline building blocks, referenced by name when (de)serializing a pipeline.
+
+| Type | Class |
+|---|---|
+| `absolute_actions_processor` | `AbsoluteActionsProcessorStep` |
+| `action_tokenizer_processor` | `ActionTokenizerProcessorStep` |
+| `add_teleop_action_as_complementary_data` | `AddTeleopActionAsComplimentaryDataStep` |
+| `add_teleop_action_as_info` | `AddTeleopEventsAsInfoStep` |
+| `device_processor` | `DeviceProcessorStep` |
+| `gripper_penalty_processor` | `GripperPenaltyProcessorStep` |
+| `gym_hil_adapter_processor` | `GymHILAdapterProcessorStep` |
+| `image_crop_resize_processor` | `ImageCropResizeProcessorStep` |
+| `intervention_action_processor` | `InterventionActionProcessorStep` |
+| `isaaclab_arena_processor` | `IsaaclabArenaProcessorStep` |
+| `libero_processor` | `LiberoProcessorStep` |
+| `map_delta_action_to_robot_action` | `MapDeltaActionToRobotActionStep` |
+| `map_tensor_to_delta_action_dict` | `MapTensorToDeltaActionDictStep` |
+| `normalizer_processor` | `NormalizerProcessorStep` |
+| `numpy2torch_action_processor` | `Numpy2TorchActionProcessorStep` |
+| `observation_processor` | `VanillaObservationProcessorStep` |
+| `policy_action_to_robot_action_processor` | `PolicyActionToRobotActionProcessorStep` |
+| `relative_actions_processor` | `RelativeActionsProcessorStep` |
+| `rename_observations_processor` | `RenameObservationsProcessorStep` |
+| `render_messages_processor` | `RenderMessagesStep` |
+| `reward_classifier_processor` | `RewardClassifierProcessorStep` |
+| `robot_action_to_policy_action_processor` | `RobotActionToPolicyActionProcessorStep` |
+| `smolvla_new_line_processor` | `NewLineTaskProcessorStep` |
+| `time_limit_processor` | `TimeLimitProcessorStep` |
+| `to_batch_processor` | `AddBatchDimensionProcessorStep` |
+| `to_batch_processor_action` | `AddBatchDimensionActionStep` |
+| `to_batch_processor_complementary_data` | `AddBatchDimensionComplementaryDataStep` |
+| `to_batch_processor_observation` | `AddBatchDimensionObservationStep` |
+| `tokenizer_processor` | `TokenizerProcessorStep` |
+| `torch2numpy_action_processor` | `Torch2NumpyActionProcessorStep` |
+| `unnormalizer_processor` | `UnnormalizerProcessorStep` |

get_embodied_tasks-0.1.0/docs/backends/lerobot/troubleshooting.md

@@ -0,0 +1,55 @@
+# LeRobot Troubleshooting & FAQ
+
+A consolidated entry point for common LeRobot problems, organized by pipeline stage.
+This is a router — each item links to the authoritative upstream page for the detail.
+Upstream docs: <https://huggingface.co/docs/lerobot>. Repo: <https://github.com/huggingface/lerobot>.
+
+> Quick triage: most "timeout" / "device not found" errors are **physical** (wiring,
+> power, USB), not software. Check hardware before code.
+
+## Installation
+- **Build fails on `av` / `ffmpeg` / native deps** → ensure `ffmpeg` is on `PATH`
+  (`lerobot-info` prints the detected version); see the upstream Installation guide.
+- **A command is "not found"** → it's behind a pip extra. See the
+  [CLI reference](./cli_reference.md) for the extra, then `pip install 'lerobot[<extra>]'`.
+- **What's my environment?** → run `lerobot-info`.
+
+## Hardware: motors, ports, power
+- **Feetech timeout / comms error (SO-100/101)** → check motor LEDs first: all steady
+  red = wiring OK; one dark = reseat the 3-pin cables / check power; blinking = error
+  state (overload, or **wrong PSU voltage** — 5V/7.4V vs 12V builds are not
+  interchangeable).
+- **Can't find the USB port** → `lerobot-find-port` (unplug when prompted); on Linux
+  `sudo chmod 666 /dev/ttyACM0` may be needed.
+- **Motors unresponsive after assembly** → `lerobot-setup-motors`, then
+  `lerobot-calibrate`; reuse the same `--robot.id` everywhere after.
+- **Damiao / CAN bus** → `lerobot-setup-can`; see the upstream Damiao page.
+
+## Cameras
+- **Camera not detected / wrong index** → `lerobot-find-cameras`; camera types are in
+  the [components reference](./components_reference.md).
+- **Policy ignores the object** → camera framing/lighting, not the model. "Can you do
+  the task from the camera view alone?" If not, fix the cameras first.
+
+## Data collection / recording
+- **Recording keys** → `→` next, `←` re-record last, `ESC` finish & upload.
+- **Fix a dataset** (delete/merge/split episodes) → `lerobot-edit-dataset`.
+- **Inspect before training** → the Hub dataset visualizer
+  (<https://huggingface.co/spaces/lerobot/visualize_dataset>).
+
+## Training
+- **Out of GPU memory** → lower `--batch_size`, use gradient accumulation, or a lighter
+  policy; see the upstream Compute Hardware Guide.
+- **LR never decays on short runs** → set `--policy.scheduler_decay_steps ≈ --steps`.
+- **Which optimizer/scheduler/policy types exist?** → [components reference](./components_reference.md).
+
+## Evaluation & benchmarks
+- **Which benchmark has my task? how to run it?** → use `get-embodied-tasks "<query>"`
+  (this repo) or the upstream benchmark docs.
+- **Benchmark env won't install locally** → use the pre-baked `docker/Dockerfile.benchmark.*`.
+- **Unstable success rate** → use `--eval.n_episodes ≥ 50`.
+
+## Still stuck?
+- Run `lerobot-info` and include its output.
+- `get-embodied-tasks doctor` checks whether the LeRobot backend is correctly located.
+- Upstream Discord: <https://discord.com/invite/s3KuuzsPFb>.