candle-org/MindAct

What it solves

MindAct provides a standardized, reproducible framework for training and evaluating imitation learning policies used in robotic desktop manipulation tasks. It eliminates the inconsistency in experiment tracking and evaluation protocols when working with different datasets and simulation environments.

How it works

The toolkit uses a protocol-based architecture with minimal adapter interfaces to integrate PyTorch, Hugging Face's LeRobot datasets and policies, and LIBERO simulation benchmarks. It manages experiments via YAML configurations and manifest-tracked provenance to ensure results can be reproduced.

Who it’s for

Researchers and developers working on embodied AI and robotic manipulation who need a consistent way to train policies and validate them against standardized benchmarks.

Highlights

• Integration: Native support for LeRobot datasets/policies and LIBERO simulations.
• Reproducibility: Uses YAML configurations and manifest tracking for experiment provenance.
• Type-Safety: Built with frozen dataclasses and runtime protocols for stability.
• Flexible Installation: Optional dependencies allow users to install only the components they need (e.g., just PyTorch or just LIBERO).
• Artifact Management: Implements a conventional directory structure for logs, checkpoints, and results.