Georgia Institute of Technology
Toward Robust, Contact-aware, and Autonomous Humanoid Loco-manipulation
Abstract
dc:description.abstractHumanoid robots are increasingly expected to operate in human-centered environments, but their deployment remains constrained by a gap between laboratory demonstrations and the reliability required in the real world. They must withstand unexpected disturbances, reason about uncertain contact, and execute long-horizon tasks with limited sensory feedback. This dissertation advances humanoid capability along two complementary dimensions: robust locomotion as a foundation for reliable operation and versatile manipulation as a source of practical value. Its central strategy is to unify locomotion and manipulation within a planner-controller hierarchy, organized around three research thrusts: robust locomotion, tactile-based whole-body manipulation, and autonomous loco-manipulation. The first thrust quantifies and enhances locomotion robustness through Linear Temporal Logic and Signal Temporal Logic. We demonstrate robust locomotion with the Cassie biped on dynamic terrain, stepping stones, and a naval ship. The second thrust extends traditional tabletop fingertip manipulation to force-aware whole-body manipulation through large-area tactile sensing. We convert human data into robot-executable motion-force plans and enable the G1 humanoid to manipulate deformable and bulky objects and to collaborate with humans during transport. The third thrust improves locomotion and loco-manipulation by fine-tuning pretrained policies with reinforcement learning. We adapt locomotion skills to slippery and uneven terrain on the Digit humanoid while preserving the structure of the pretrained behavior. We also improve loco-manipulation success and controller tracking accuracy on the Booster T1 humanoid, with less demonstration data than conventional behavior cloning. Together, these contributions establish a progression from formally grounded locomotion robustness to contact-aware whole-body manipulation and autonomous loco-manipulation. The dissertation shows how formal models, tactile demonstration learning, and reinforcement learning fine-tuning advance humanoid robots toward stable locomotion, autonomous manipulation, and rich physical interaction.
Degree
thesis:*- Name thesis:degree_name
- Robotics, PhD
- Department thesis:degree_department
- Mechanical Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gu, Zhaoyuan
- Advisor dc:contributor.advisor
-
- Zhao, Ye
- Committee members dc:contributor.committeemember
-
- Ha, Sehoon
- Tucker, Maegan
- Shi, Guanya
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82159