Georgia Institute of Technology
Cross-Domain Imitation Learning from Human Videos for Robotic Manipulation
Abstract
dc:description.abstractLearning robotic manipulation from human videos offers a scalable alternative to costly robot-specific data collection, yet a substantial domain gap across visual and kinematic dimensions hinders direct transfer. This thesis presents ImMimic, an embodiment-agnostic co-training framework that bridges this gap by combining Dynamic Time Warping trajectory mapping with MixUp interpolation in latent and action space. Evaluated on four real-world manipulation tasks across four robotic embodiments, ImMimic consistently improves task success rates and execution smoothness over robot-only and standard co-training baselines.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, MS
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shin, Woo Chul
- Advisor dc:contributor.advisor
-
- Xu, Danfei
- Committee members dc:contributor.committeemember
-
- Kousik, Shreyas
- Ravichandar, Harish
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81668