Learning Dexterous Manipulation from Grasp Priors and Force-Aware Demonstrations - Robotics Institute Carnegie Mellon University
Loading Events

MSR Thesis Presentation

November

19
Thu
Ying Yuan PhD Student Robotics Institute,
Carnegie Mellon University
Thursday, November 19
11:00 am to 12:00 pm
GHC 4405
Learning Dexterous Manipulation from Grasp Priors and Force-Aware Demonstrations
Abstract:
Dexterous manipulation requires coordinated finger motion, stable hand–object contact, and appropriate force application. Learning these behaviors from demonstrations is challenging because task-specific robot data is costly to collect and motion commands alone do not always produce reliable contact. This thesis studies two complementary approaches to improving the data used for dexterous policy learning: large-scale grasp pretraining and force-aware demonstration collection.
First, we present a hierarchical imitation learning framework that combines high-level hand sub-goal prediction with a low-level controller pretrained on grasp trajectories. We construct G2D-Pretrain from large-scale grasp annotations and introduce DexCraft, a six-task benchmark for articulated tool use. Simulation and real-world experiments show that grasp pretraining improves functional manipulation beyond policies trained from task demonstrations alone. Second, we present Force-Modulated Dexterity (FOMO-Dex), a unified compliance formulation that uses robot-side force feedback to adjust selected components of nominal arm and finger commands. Across four contact-rich tasks, FOMO-Dex improves contact behavior, supports successful demonstration collection, and provides corrected action targets for behavior cloning. Together, these studies show how broad grasp experience and force-aware task demonstrations can support learning contact-rich dexterous manipulation.

Thesis Committee:
David Held (chair)
Guanya Shi
Shubham Tulsiani
Eliot Xing