SimDuoReal: Object-Centric Bimanual Policy for Emergent Dexterity
Abstract
Bimanual dexterous manipulation requires coordinated arm and hand control, yet specifying such coordination manually is challenging and often task-specific. In this work, we explore whether complex bimanual dexterity can emerge from a simple object-centric learning objective. We train reinforcement learning policies under a shared object-centric reward, varying their training environments rather than tuning separate rewards for each behavior. We find that the policies discover pre-grasp manipulation, grasping and regrasping, in-hand adjustment and handover to reach object-pose goals without hand-designed motion sequences. We then refine the experts on transition states and use DAgger and reinforcement learning to distill them into a single policy that composes these behaviors for diverse 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 dexterity, shifting the burden from manually designing robot coordination to specifying what should happen to the object.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper presents a novel object-centric RL approach for bimanual dexterity that successfully transfers zero-shot to real hardware. It demonstrates that complex coordination behaviors can emerge from simple object-goal rewards, offering a scalable and generalizable framework for dexterous manipulation that outperforms single-arm baselines.
Methodology
The paper proposes an object-centric reinforcement learning framework for bimanual dexterity. Instead of manually designing complex coordination rewards for arms and hands, the authors utilize a shared object-centric reward that drives the policy toward desired object configurations. The methodology involves training expert policies in varied simulated environments to discover emergent behaviors (pre-grasp, in-hand manipulation, handover) and then distilling these experts into a single generalist policy using DAgger and RL. This approach shifts the design burden from kinematic coordination to task-level object goals, which is a significant conceptual simplification for dexterous manipulation.
Results & evaluation
The evaluation is strong, featuring both simulation and real-world hardware tests. In simulation, the policy achieves 85.0% success in barcode scanning, outperforming single-arm baselines. On real hardware, the policy demonstrates zero-shot transfer from simulation, achieving a 76.3% success rate across four diverse tasks (assembly, scanning, cleaning, cooking) with unseen objects. The inclusion of real-world results with zero-shot transfer is a major strength, as sim-to-real gaps are a persistent challenge in dexterous robotics.
Reproducibility
The paper provides a project URL and mentions the use of standard techniques like DAgger and RL. However, specific details on the simulation environment setup, reward function formulations, and hyperparameters for the distillation process are likely in the appendix (not fully visible in the provided text snippet but implied). The reliance on specific hardware (dexterous arm-hand systems) may limit immediate reproducibility for labs without similar setups, but the core RL methodology is standard.
Limitations
The success rate on real hardware (76.3%) indicates a significant failure rate, suggesting that while the approach is promising, it is not yet robust enough for fully autonomous deployment in unstructured environments. The paper focuses on a specific set of tasks; generalization to entirely different object geometries or dynamic environments is not fully explored. The computational cost of training multiple experts and distilling them may be high.
Broader impact
This work contributes to the broader goal of general-purpose robot manipulation. By demonstrating that complex bimanual coordination can emerge from simple object-centric objectives, it offers a scalable path toward dexterous robots that do not require hand-coded motion primitives for every new task. This could accelerate the development of general-purpose service robots and industrial automation systems. The paper presents a novel object-centric RL approach for bimanual dexterity that successfully transfers zero-shot to real hardware. It demonstrates that complex coordination behaviors can emerge from simple object-goal rewards, offering a scalable and generalizable framework for dexterous manipulation that outperforms single-arm baselines.
Author of this paper? Send a correction.