act-plus-plus
Routes ACT++, ACT, Diffusion Policy, VINN, and MuJoCo simulation workflows for bimanual ALOHA episode data and imitation-learning tasks.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- act-plus-plus
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes ACT++, ACT, Diffusion Policy, VINN, and MuJoCo simulation workflows for bimanual ALOHA episode data and imitation-learning tasks.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/api-reference.md5.8 KB
- 参考文档references/cli-reference.md4.7 KB
- 参考文档references/data-formats.md3.9 KB
- 参考文档references/overview.md3.6 KB
- 参考文档references/repo-provenance.md2.3 KB
- 参考文档references/repo-routing-metadata.json520 B
- 参考文档references/troubleshooting.md5.6 KB
03 SKILL.md 正文
act-plus-plus
Use this repo skill when a task is about the ACT++ / Mobile ALOHA imitation-learning repository: simulated ALOHA episode generation, ACT or Diffusion Policy training, HDF5 episode utilities, MuJoCo + DM Control task behavior, or VINN feature workflows.
Start here
- Read repo provenance if you need to check whether this skill matches the checkout or if code changed.
- Run or read scripts/check_environment.py before expensive workflows. It checks the external package/backend stack without running training.
- Use the route table below, then stay inside the owning sub-skill and its linked references.
- For command flags and data shapes that cross multiple workflows, read CLI reference, data formats, and API reference.
- For backend, dependency, or real-robot caveats, read cross-cutting troubleshooting.
Install notes
For a checkout, install the repo's Python stack before launching any workflow:
- install the external Python dependencies needed by the workflow you want;
- install the repository itself in editable mode;
- install the
detr/subpackage in editable mode so theutilandmodelsimports resolve; - then run
scripts/check_environment.pybefore long jobs.
The exact dependency set depends on the chosen workflow: simulation only needs MuJoCo / DM Control plus image and HDF5 utilities, while policy training and VINN also need CUDA-enabled torch plus robomimic / diffusers support.
Route by user intent
| User task | Read |
|---|---|
| Generate scripted simulated episodes, replay actions, render videos, mirror/compress/truncate HDF5 episodes, or debug MuJoCo/DM Control rendering | simulation-data |
| Train/evaluate ACT, CNNMLP, Diffusion Policy, or the ACT VQ latent model; inspect checkpoint/stat files; convert README commands into current CLI flags | policy-training |
Cache BYOL/ResNet image features, select nearest-neighbor k, inspect VINN feature-file layouts, or avoid the raw VINN scripts' interactive traps |
vinn-offline |
| Understand supported task names, camera names, HDF5 schemas, or how episode files flow between data and training | overview and data formats |
Diagnose ModuleNotFoundError, BOX_POSE assertion failures, missing MUJOCO_GL, CUDA errors, mismatched robomimic/diffusers installs, or Mobile ALOHA hardware dependencies |
troubleshooting |
Repository operating model
- Core simulated tasks are transfer cube and bimanual insertion. The sim data flow uses an end-effector environment for scripted demos, then replays joint commands into the joint-space environment to record observations.
- Episode files are HDF5 datasets with
/observations/qpos,/observations/qvel,/action, and/observations/images/<camera>groups. Compressed/mirrored variants addcompress=trueand/compress_len. - ACT/CNNMLP/Diffusion training is driven by
imitate_episodes-style arguments. Current training code uses--num_steps; older README command snippets may say--num_epochs. - Training and VINN paths call
.cuda()directly. Treat CUDA as required unless the code is explicitly modified. - MuJoCo rendering needs an offscreen GL backend such as
MUJOCO_GL=egl; do not treat a CPU import as proof that rendered sim data generation works. - Real robot branches depend on the external Mobile ALOHA / Interbotix stack and hardware. This generated skill covers them only as troubleshooting/gap notes, not as a verified runtime route.
Minimal checks
Use the bundled checkers rather than launching long training or real robot code first:
python scripts/check_environment.py
python sub-skills/simulation-data/scripts/check_sim_backend.py --repo-root /path/to/act-plus-plus --task sim_transfer_cube
python sub-skills/policy-training/scripts/check_policy_stack.py --repo-root /path/to/act-plus-plus
python sub-skills/vinn-offline/scripts/check_vinn_stack.py --repo-root /path/to/act-plus-plus
The --repo-root examples accept any checkout of ACT++; they are not tied to the source checkout used to build this skill.
What is intentionally not routed
- Servo calibration,
align-style robot movement, Dynamixel diagnostics, and real Mobile ALOHA runtime deployment require external hardware packages and are not verified here. train_actuator_networkis an experiment-oriented utility with hard-coded data/log paths and dataset fields; treat it as reference evidence only.- The empty
byol_pytorchgitlink in the inspected checkout was not used as source evidence. VINN guidance is based on the available ACT++ VINN scripts and feature-file contracts.
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.