SimDuoReal: Object-Centric Bimanual Policy for Emergent Dexterity
SimDuoReal: Object-Centric Bimanual Policy for Emergent Dexterity
This research introduces SimDuoReal, a reinforcement learning framework that achieves emergent bimanual dexterity through an object-centric reward objective. By training policies in varying environments rather than tuning task-specific rewards, the system discovers behaviors such as pre-grasp manipulation, grasping, regrasping, in-hand adjustment, and handover without manual motion sequences. The distilled single policy successfully completes 61 of 80 trials (76.3%) on real hardware across four tasks with unseen objects and no fine-tuning. In simulated barcode scanning, it achieves an 85.0% success rate compared to 60.4% for two single-arm policies without handover. The work suggests that specifying object-level goals provides a scalable interface for learning general bimanual manipulation, shifting the burden from manual coordination design to defining desired object interactions.
The study demonstrates that a shared object-centric reward function enables reinforcement learning policies to discover complex bimanual behaviors like pre-grasp manipulation and handover without manual motion design, achieving 76.3% success on real hardware across four tasks with unseen objects.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
trials
80 trials
Reported trials
View original evidence
iverse bimanual tasks. On real hardware, the same policy succeeds in 61 of 80 trials (76.3\%) across four tasks with unseen objects and no task-specific fine-tuning. In simulated barcode scanning, our policy achieves 85.0\% success, compared with 60.4\% for two single-arm policies without handover. These results suggest that this objective provides a simple and scalable interface for learning general bimanual dexteri
Open source S4robots
2
View original evidence
iverse bimanual tasks. On real hardware, the same policy succeeds in 61 of 80 trials (76.3\%) across four tasks with unseen objects and no task-specific fine-tuning. In simulated barcode scanning, our policy achieves 85.0\% success, compared with 60.4\% for two single-arm policies without handover. These results suggest that this objective provides a simple and scalable interface for learning general bimanual dexteri
Open source S4tasks
4
View original evidence
iverse bimanual tasks. On real hardware, the same policy succeeds in 61 of 80 trials (76.3\%) across four tasks with unseen objects and no task-specific fine-tuning. In simulated barcode scanning, our policy achieves 85.0\% success, compared with 60.4\% for two single-arm policies without handover. These results suggest that this objective provides a simple and scalable interface for learning general bimanual dexteri
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
iverse bimanual tasks. On real hardware, the same policy succeeds in 61 of 80 trials (76.3\%) across four tasks with unseen objects and no task-specific fine-tuning. In simulated barcode scanning, our policy achieves 85.0\% success, compared with 60.4\% for two single-arm policies without handover. These results suggest that this objective provides a simple and scalable interface for learning general bimanual dexteri
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org