Skip to content

AGENTS.md

Latest commit

聽

History

History
26 lines (18 loc) 路 1.16 KB

AGENTS.md

File metadata and controls

26 lines (18 loc) 路 1.16 KB
Raw

MindAct agent guidance

Scope

MindAct is a PyTorch- and Hugging Face-native toolkit for reproducible embodied-policy training and evaluation. The active package lives under src/mindact; historical MindNLP and MindTorch code is preserved on the legacy branch.

Engineering rules

  • Keep core imports lightweight. Never import torch, lerobot, or libero at module import time; use mindact.utils.imports.require_module() inside integration entry points.
  • Use @runtime_checkable protocols for integration boundaries instead of ABCs.
  • Use frozen, keyword-only dataclasses for configuration and result records.
  • Reject unknown configuration fields and validate values at construction time.
  • Preserve experiment provenance in JSON-compatible manifests.
  • Do not copy or reimplement upstream LeRobot or LIBERO internals in the core package.

Validation

From a Python 3.12+ environment with development dependencies installed:

pytest -q
ruff check .
python -m build

Optional integration tests must skip cleanly when their dependencies are not installed. Do not commit checkpoints, rollout media, trajectory data, simulator caches, or experiment outputs.