lerobot
Build and debug LeRobot datasets, training, and policy evaluation. For a first pretrained robot-arm demo, start with architect's reference-app selection.
Install
npx skills add https://github.com/robium-ai/robium/tree/main/skills/lerobot
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install robium-ai-robium@llmmart
git clone https://github.com/robium-ai/robium.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole robium-ai/robium collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
LeRobot
Follow one contract chain through LeRobot: embodiment, dataset, processor, checkpoint, runtime observations and actions, then evaluation. Find the first contract that does not match.
Establish the contract
- For a new manipulation app or first pretrained-policy demo, read architect before creating an environment or training pipeline. It finds the saved apps checkout and selects a compatible baseline. If selection already happened, continue here. Existing dataset, training, or evaluation work does not need onboarding; an inference-only request does not authorize training.
- Inspect the installed LeRobot version and current CLI help before writing flags. Dataset formats, policy families, scripts, and extras change quickly.
- Match the robot's state, action space, cameras, rates, and task to the dataset. A policy adapts to those features; it cannot repair a mismatched embodiment.
- Inspect every checkpoint's configuration and processor files, not only its weights. Base and fine-tuned checkpoints from one family can expect different camera layouts.
- Keep Hub identity, transfer, publication, and Jobs lifecycle at the Hugging Face boundary. Keep source-selection strategy in data.
Prove the loop cheaply
- Start with a small shipped policy and a known dataset/environment pair.
- Run a short train that writes a checkpoint, then load that exact checkpoint through evaluation. Completion and numeric metrics are the smoke-test result; policy quality is not.
- Confirm loss, saved processors, input/output shapes, rollout metrics, and video or real-robot behavior before increasing steps or hardware cost.
- Treat real-hardware rollout as a new safety boundary even when simulation evaluation passed.
Go deeper only when needed
- For loading, recording, editing, migration, and episode visualization, read references/datasets.md.
- For policy choice, camera remapping, training, compute sizing, and remote Jobs behavior, read references/policies-and-training.md.
- For simulation evaluation, headless rendering, EnvHub, or real-hardware rollout, read references/eval-and-sim.md.
- When a checkpoint, feature contract, evaluation worker, dependency, or remote run fails, start with FAILURES.md.
- The concrete PushT examples are useful only when that smoke path matches the application: load dataset and train ACT.
- Use the current LeRobot documentation and source for version-sensitive APIs and the shipped policy/environment list.
Done
- The target dataset loads, the checkpoint carries its processors, evaluation exercises matching observations and actions, and measured results justify any longer run or hardware deployment.
Files (robium)
-
examples
-
load-dataset-snippet.py 2.4 KB
Loading… -
train-act-command.md 4.2 KB
Loading…
-
-
references
-
datasets.md 9.4 KB
Loading… -
eval-and-sim.md 7.3 KB
Loading… -
policies-and-training.md 5 KB
Loading…
-
-
evals.yaml 555 B
Loading… -
FAILURES.md 3.3 KB
Loading… -
SKILL.md 3.1 KB
--- name: lerobot description: Build and debug LeRobot datasets, training, and policy evaluation. For a first pretrained robot-arm demo, start with architect's reference-app selection. --- # LeRobot Follow one contract chain through LeRobot: embodiment, dataset, processor, checkpoint, runtime observations and actions, then evaluation. Find the first contract that does not match. ## Establish the contract - For a new manipulation app or first pretrained-policy demo, read [architect](../architect/SKILL.md) before creating an environment or training pipeline. It finds the saved apps checkout and selects a compatible baseline. If selection already happened, continue here. Existing dataset, training, or evaluation work does not need onboarding; an inference-only request does not authorize training. - Inspect the installed LeRobot version and current CLI help before writing flags. Dataset formats, policy families, scripts, and extras change quickly. - Match the robot's state, action space, cameras, rates, and task to the dataset. A policy adapts to those features; it cannot repair a mismatched embodiment. - Inspect every checkpoint's configuration and processor files, not only its weights. Base and fine-tuned checkpoints from one family can expect different camera layouts. - Keep Hub identity, transfer, publication, and Jobs lifecycle at the Hugging Face boundary. Keep source-selection strategy in data. ## Prove the loop cheaply - Start with a small shipped policy and a known dataset/environment pair. - Run a short train that writes a checkpoint, then load that exact checkpoint through evaluation. Completion and numeric metrics are the smoke-test result; policy quality is not. - Confirm loss, saved processors, input/output shapes, rollout metrics, and video or real-robot behavior before increasing steps or hardware cost. - Treat real-hardware rollout as a new safety boundary even when simulation evaluation passed. ## Go deeper only when needed - For loading, recording, editing, migration, and episode visualization, read [references/datasets.md](references/datasets.md). - For policy choice, camera remapping, training, compute sizing, and remote Jobs behavior, read [references/policies-and-training.md](references/policies-and-training.md). - For simulation evaluation, headless rendering, EnvHub, or real-hardware rollout, read [references/eval-and-sim.md](references/eval-and-sim.md). - When a checkpoint, feature contract, evaluation worker, dependency, or remote run fails, start with [FAILURES.md](FAILURES.md). - The concrete PushT examples are useful only when that smoke path matches the application: [load dataset](examples/load-dataset-snippet.py) and [train ACT](examples/train-act-command.md). - Use the current [LeRobot documentation](https://huggingface.co/docs/lerobot) and [source](https://github.com/huggingface/lerobot) for version-sensitive APIs and the shipped policy/environment list. ## Done - The target dataset loads, the checkpoint carries its processors, evaluation exercises matching observations and actions, and measured results justify any longer run or hardware deployment.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.