Learning goal-conditioned in-hand object re-orientation

Dexterous manipulation using anthropomorphic hands is a critical capability for general-purpose robots operating in human-centric, unstructured environments. However, controlling a dexterous hand to perform contact-rich tasks remains challenging, due to its high degrees of freedom and the need to maintain force-closure. While classical model-based methods have shown promising results, their applicability is often constrained by strong assumptions that limit generalization. In this thesis, we address the problem of goal-conditioned in-hand object re-orientation over the full SO(3) space, in contrast to prior works that restrict rotation to specific axes. We propose a learning-based framework that combines model-free reinforcement learning with a teacher-student paradigm and automatic domain randomization to train a unified policy for in-hand object rotation. Our policy relies only on proprioceptive feedback and object pose tracking, without requiring prior knowledge of object shape. We evaluate the approach extensively in simulation, demonstrating strong generalization to objects with unseen shapes and physical properties. Furthermore, we validate the policy on a low-cost, open-source dexterous hand, achieving successful zero-shot transfer to the real world.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Niu, Yilong

Contributors dc:contributor

• Yuan, Wenzhen

• Dexterous Manipulation
• Reinforcement Learning

Statement dc:rights

• Copyright 2025 Yilong Niu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130228

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130228