diffusion-policy
Use Diffusion Policy for robot imitation-learning configs, zarr replay data, policy/model interfaces, training/evaluation workflows, and safety-gated real robot operations.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- diffusion-policy
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use Diffusion Policy for robot imitation-learning configs, zarr replay data, policy/model interfaces, training/evaluation workflows, and safety-gated real robot operations.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/overview-and-installation.md4.2 KB
- 参考文档references/repo-provenance.md2.4 KB
- 参考文档references/repo-routing-metadata.json532 B
- 参考文档references/troubleshooting.md4.6 KB
03 SKILL.md 正文
Diffusion Policy
Use this repo skill when a task involves the diffusion_policy robotics imitation-learning codebase: Hydra experiment configs, Push-T/Robomimic/Kitchen/BlockPush tasks, zarr demonstration replay buffers, diffusion action policies, checkpoint evaluation, Ray multiruns, or real Push-T robot data collection/evaluation.
Start with these checks
-
Read repo provenance before assuming the skill matches a checkout or package version.
-
Read overview and installation for dependency variants, package/import behavior, and safe smoke checks.
-
Run the bundled smoke checker when you need environment evidence without training, downloads, Ray, cameras, or robot motion:
python scripts/smoke_check.py --json -
Use cross-cutting troubleshooting for import, Hydra, dependency, dataset, CUDA, simulator, W&B, and hardware failures.
Route by task
| If the user needs... | Read |
|---|---|
| Build or debug single-seed training, checkpoint evaluation, Hydra overrides, Ray multiruns, output trees, W&B/logs, or benchmark command structure | training-and-evaluation |
Inspect or validate zarr/zip ReplayBuffer stores, dataset sample schemas, SequenceSampler, normalizers, or dataset conversion assumptions |
data-and-replay-buffers |
| Choose/inspect low-dim, image, hybrid, diffusion UNet/Transformer, Robomimic, BET, or IBC policy/model families; debug shape, normalizer, checkpoint, or device issues | policies-and-models |
| Preflight UR5 + RealSense + SpaceMouse workflows, demo capture, real robot policy evaluation, shared-memory IO, real dataset conversion, or hardware safety gates | real-robot-operations |
What this skill covers
- The core
Dataset->Normalizer->Policy->EnvRunner->Workspaceworkflow. - Low-dimensional and image observation/action interfaces, including horizon terminology (
horizon,n_obs_steps,n_action_steps). - Hydra workspace/task config composition and command-building patterns.
- ReplayBuffer zarr/zip store structure and episode validation.
- Checkpoint evaluation behavior, including EMA model selection.
- Optional simulator, CUDA, Ray, W&B, Robomimic, MuJoCo, and real-robot dependency boundaries.
- Safety-gated real-robot operations: UR RTDE, RealSense, SpaceMouse, shared-memory queues/ring buffers, timestamp alignment, and conversion of recorded episodes.
Boundaries and safety
- Do not launch training, Ray workers, data downloads, W&B online logging, simulator rollouts, camera capture, or robot motion as a smoke check.
- Do not treat a CPU import check as proof that benchmark-scale CUDA, MuJoCo/Robomimic, or real-robot workflows are ready.
- Stop for explicit operator confirmation before any live robot action, camera recording, or command that may overwrite run outputs.
- When a workflow needs project entrypoints or config files, first verify that the user is operating in a compatible Diffusion Policy checkout or equivalent project layout; this skill provides reusable operating knowledge and safe helpers, not the full benchmark runtime.
Bundled scripts
- scripts/smoke_check.py checks distribution metadata, representative imports, optional config root counts, and optional torch CUDA visibility without side effects.
- Sub-skill helpers validate config composition, summarize multirun logs, inspect ReplayBuffer stores, inspect policy interfaces, and preflight real-robot dependencies.
Refresh and verification notes
This skill was generated from a clean source snapshot before skill-output files were added. If a checkout has a different commit, changed config targets, dependency files, entrypoint options, dataset/policy/workspace classes, or real-robot APIs, refresh the repo skill before relying on detailed commands or signatures.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.