LodeStar: Long-horizon Dexterity via Synthetic Data Augmentation from Human Demonstrations

TL;DR

LodeStar introduces a framework that decomposes human demonstrations into skills, generates synthetic data via reinforcement learning, and trains a transformer-based policy to achieve robust long-horizon robotic manipulation.

Contribution

The paper presents a novel system combining off-the-shelf foundation models and reinforcement learning to augment datasets and improve long-horizon dexterous manipulation.

Findings

Significant performance improvements over baselines.

Enhanced robustness in real-world tasks.

Effective skill chaining with Skill Routing Transformer.

Abstract

Developing robotic systems capable of robustly executing long-horizon manipulation tasks with human-level dexterity is challenging, as such tasks require both physical dexterity and seamless sequencing of manipulation skills while robustly handling environment variations. While imitation learning offers a promising approach, acquiring comprehensive datasets is resource-intensive. In this work, we propose a learning framework and system LodeStar that automatically decomposes task demonstrations into semantically meaningful skills using off-the-shelf foundation models, and generates diverse synthetic demonstration datasets from a few human demos through reinforcement learning. These sim-augmented datasets enable robust skill training, with a Skill Routing Transformer (SRT) policy effectively chaining the learned skills together to execute complex long-horizon manipulation tasks. Experimental evaluations on three challenging real-world long-horizon dexterous manipulation tasks demonstrate that our approach significantly improves task performance and robustness compared to previous baselines. Videos are available at lodestar-robot.github.io.