MindAct

A PyTorch and Hugging Face native training and evaluation toolkit for embodied policies

MindAct is a reproducible training and evaluation framework for imitation learning policies on desktop manipulation tasks. It integrates LeRobot datasets and policies with LIBERO simulation benchmarks, providing experiment provenance tracking and standardized evaluation protocols.

Features

• Protocol-based architecture: Minimal adapter interfaces for datasets, policies, and environments
• Reproducible experiments: YAML configurations with manifest-tracked provenance
• Optional dependencies: Lazy loading of torch, lerobot, and libero via extras
• Artifact management: Conventional directory structure for checkpoints, logs, and results
• Type-safe: Frozen dataclasses and runtime protocols throughout

Quick Start

Installation

# Basic installation (configuration and CLI only)
pip install -e .

# With PyTorch
pip install -e ".[torch]"

# With LeRobot datasets and policies
pip install -e ".[lerobot]"

# With LIBERO simulation environments
pip install -e ".[libero]"

# All integrations
pip install -e ".[torch,lerobot,libero]"

# Development tools
pip install -e ".[dev]"

Run an Example

# Validate configuration
mindact config-check configs/experiments/libero-baseline.yaml

# Run the dependency-free evaluation smoke path
mindact eval configs/experiments/libero-baseline.yaml \
  --runner fake \
  --run-id smoke-run \
  --output-dir outputs \
  --episodes 2

config-check validates a configuration, and eval --runner fake exercises the full evaluation lifecycle with built-in test doubles. Real LeRobot policies and LIBERO environments arrive with their adapter implementations; the fake runner is a contract smoke test, not a benchmark result.

Configuration Format

name: libero-baseline
seed: 42
output_dir: outputs

dataset:
  repo_id: lerobot/aloha_sim_insertion_human
  revision: null
  split: train

policy:
  name: act
  pretrained_model: null

environment:
  name: libero
  task_suite: libero_spatial
  task_ids: []

training:
  steps: 10000
  batch_size: 8
  learning_rate: 0.0001
  log_every: 100
  checkpoint_every: 2000

evaluation:
  episodes: 20
  max_steps: 500
  record_video: false

Project Status

MindAct v0.1 is the initial skeleton release. Core interfaces and configuration system are stable. Implementation priorities:

Phase B (current): Unified reproducible training and evaluation
Phase A (future): Trajectory quality diagnostics

See docs/ for architecture details and contribution guidelines.

Repository History

This repository was originally MindNLP, a MindSpore-based NLP library. The legacy codebase is preserved in the legacy branch. MindAct represents a complete pivot to embodied AI with PyTorch and Hugging Face as the native stack.

License

Apache License 2.0. See LICENSE and NOTICE for details.