Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Wan, Weikang, Fu, Jiawei, Yuan, Xiaodi, Zhu, Yifeng, Su, Hao
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2508.17547
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916914745311232
author Wan, Weikang
Fu, Jiawei
Yuan, Xiaodi
Zhu, Yifeng
Su, Hao
author_facet Wan, Weikang
Fu, Jiawei
Yuan, Xiaodi
Zhu, Yifeng
Su, Hao
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2508_17547
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle LodeStar: Long-horizon Dexterity via Synthetic Data Augmentation from Human Demonstrations
Wan, Weikang
Fu, Jiawei
Yuan, Xiaodi
Zhu, Yifeng
Su, Hao
Robotics
Artificial Intelligence
Machine Learning
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.
title LodeStar: Long-horizon Dexterity via Synthetic Data Augmentation from Human Demonstrations
topic Robotics
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2508.17547