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.
- 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
# 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]"# 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 2config-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.
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: falseMindAct 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.
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.
Apache License 2.0. See LICENSE and NOTICE for details.
