Bimanual Rope Manipulation Policy Learned from Human Teleoperation Data
A recent study on arXiv explores imitation learning applied to the bimanual handling of deformable linear objects (DLOs), such as ropes and cables. The researchers evaluated two policies utilizing Action Chunking with Transformers, both trained on identical teleoperation data. One policy is vision-based, relying on two egocentric RGB streams captured by cameras mounted on the wrists, while the other is state-based, utilizing the 3D coordinates of the DLO. This research tackles the complexities of DLO manipulation, including the challenges posed by infinite-dimensional configuration spaces and self-occlusion, with the goal of enhancing generalization from limited datasets.
Key facts
- arXiv paper 2605.16043
- Focuses on bimanual rope manipulation
- Uses imitation learning from teleoperation
- Compares vision-based and state-based policies
- Action Chunking with Transformers architecture
- Egocentric RGB streams from wrist cameras
- State-based policy uses 3D points of DLO
- Addresses generalization from small datasets
Entities
Institutions
- arXiv