Reproduce “PRIME: Scaffolding Manipulation Tasks with Behavior Primitives for Data-Efficient Imitation Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tian Gao, Soroush Nasiriany, Huihan Liu, Quantao Yang, Yuke Zhu

PRIME is a framework that enhances data efficiency in robot imitation learning by scaffolding manipulation tasks with behavior primitives. It automatically parses raw demonstrations into primitive sequences and trains a high-level policy, achieving 10-34% higher success rates in simulation and 20-48% higher rates on real robots compared to state-of-the-art baselines using only 30 human demonstrations per task.

Key insights

  • Decomposing complex, long-horizon tasks into sequences of simpler, robust behavior primitives significantly reduces the effective temporal horizon for imitation learning, leading to greater data efficiency.
  • A self-supervised Inverse Dynamics Model, trained on task-agnostic primitive rollouts, can effectively parse raw human demonstrations into primitive sequences, eliminating the need for manual segmentation labels.
  • Pretraining the high-level policy with diverse primitive rollouts and employing stepwise data augmentation on segmented primitive sequences are crucial for enhancing data efficiency and policy generalization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.