DexCtrl: Sim-To-Real Dexterity with Adaptive Controller Learning
Shuqi Zhao, Ke Yang, Yuxin Chen, Chenran Li, Yichen Xie, Xiang Zhang, Changhao Wang, Masayoshi Tomizuka
AI summary
Problem
Discrepancies between simulation and real-world low-level controller dynamics cause performance drops during policy transfer. Current solutions rely on labor-intensive manual tuning or randomization that can increase training difficulty.
Approach
A framework that jointly learns desired actions and adaptive controller parameters by leveraging historical trajectory and control data via a student policy distilled from an RL oracle.
Key results
- Significantly outperforms baselines in zero-shot sim-to-real transfer for rotation and flipping tasks
- Improved stability (Time To Fail) and speed (Rotation Reward) over fixed-controller methods
- Reduces the need for manual tuning of controller parameters
- Demonstrates robustness across real-world objects with varying masses and frictions
Why it matters
It enables more reliable transfer of complex, contact-rich robotic manipulation policies from simulation to reality with minimal human intervention.
Abstract
Dexterous manipulation has advanced rapidly, with policies now capable of performing complex, contact-rich tasks in simulation. However, transferring these policies from simulation to real world remains a significant challenge. A key obstacle is the mismatch in low-level controller dynamics, where same trajectories can produce vastly different contact forces and behaviors when control parameters change. Existing solutions often rely on manual tuning or controller randomization, which can be labor-intensive, task-specific, and introduce substantial training difficulty. In this work, we propose DexCtrl, a novel framework that jointly learns actions and controller parameters by leveraging the historical information of both trajectory and controller. This adaptive controller adjustment mechanism enables the policy to automatically tune control parameters during execution, thereby mitigating severe sim-to-real gap without extensive manual tuning or excessive randomization. Moreover, by explicitly providing controller parameters as part of the observation, our approach facilitates better reasoning over force interactions and improves robustness in real-world scenarios. Experimental results demonstrate that our method achieves improved transfer performance across a variety of dexterous tasks involving variable force conditions.