Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DexCtrl: Towards Sim-to-Real Dexterity with Adaptive Controller Learning

Shuqi Zhao, Ke Yang, Yuxin Chen, Chenran Li, Yichen Xie, Xiang Zhang, Changhao Wang, Masayoshi Tomizuka

TL;DR

DexCtrl addresses the critical sim-to-real challenge in dexterous manipulation by explicitly accounting for controller dynamics. It introduces a dual-module framework that jointly outputs actions and controller parameters, guided by historical trajectory information, to adapt control in real time. Through PPO-trained oracle data and subsequent distillation into action- and parameter-predictor modules, the method improves transfer performance in rotation and flipping tasks across simulation and real hardware, particularly when controller adaptation is enabled. The work highlights how learned controller parameters encode force requirements and demonstrates potential for more robust, hardware-aware dexterous manipulation with limited manual tuning.

Abstract

Dexterous manipulation has seen remarkable progress in recent years, with policies capable of executing many complex and contact-rich tasks in simulation. However, transferring these policies from simulation to real world remains a significant challenge. One important issue is the mismatch in low-level controller dynamics, where identical trajectories can lead to vastly different contact forces and behaviors when control parameters vary. Existing approaches often rely on manual tuning or controller randomization, which can be labor-intensive, task-specific, and introduce significant training difficulty. In this work, we propose a framework that jointly learns actions and controller parameters based on the historical information of both trajectory and controller. This adaptive controller adjustment mechanism allows the policy to automatically tune control parameters during execution, thereby mitigating the 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.

DexCtrl: Towards Sim-to-Real Dexterity with Adaptive Controller Learning

TL;DR

DexCtrl addresses the critical sim-to-real challenge in dexterous manipulation by explicitly accounting for controller dynamics. It introduces a dual-module framework that jointly outputs actions and controller parameters, guided by historical trajectory information, to adapt control in real time. Through PPO-trained oracle data and subsequent distillation into action- and parameter-predictor modules, the method improves transfer performance in rotation and flipping tasks across simulation and real hardware, particularly when controller adaptation is enabled. The work highlights how learned controller parameters encode force requirements and demonstrates potential for more robust, hardware-aware dexterous manipulation with limited manual tuning.

Abstract

Dexterous manipulation has seen remarkable progress in recent years, with policies capable of executing many complex and contact-rich tasks in simulation. However, transferring these policies from simulation to real world remains a significant challenge. One important issue is the mismatch in low-level controller dynamics, where identical trajectories can lead to vastly different contact forces and behaviors when control parameters vary. Existing approaches often rely on manual tuning or controller randomization, which can be labor-intensive, task-specific, and introduce significant training difficulty. In this work, we propose a framework that jointly learns actions and controller parameters based on the historical information of both trajectory and controller. This adaptive controller adjustment mechanism allows the policy to automatically tune control parameters during execution, thereby mitigating the 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.
Paper Structure (14 sections, 2 equations, 7 figures, 4 tables)

This paper contains 14 sections, 2 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Problem Statements
  3. Methods
  4. Oracle Policy for Data Collection
  5. Action Prediction and Control Parameters Prediction
  6. Experiments
  7. Experimental Setup
  8. Does our Method Improve the Original Oracle Performance?
  9. Does our Method Narrow the Sim-to-real Gap?
  10. How does our Method Perform across Objects with Varying Physical Parameters?
  11. How do Changes in Controller Parameters Impact the Results?
  12. Related Work
  13. Conclusion
  14. Limitations

Figures (7)

  • Figure 1: Compared to previous work with only action prediction (upper left), DexCtrl (lower left) jointly predicts both action and control parameters, significantly reducing human labor for tuning and achieving better performance on two contact-rich manipulation tasks: rotation and flipping.
  • Figure 2: Overview framework of DexCtrl, where $\hat{a}_t$ and $\hat{K}_t$ mean predicted joint position actions and predicted control parameters, respectively.
  • Figure 3: Flipping task performance in simulation and real world.
  • Figure 4: Real-world results of object rotation with different physical parameters.
  • Figure 5: Visualization for same-shape objects rotation with different masses and frictions.
  • ...and 2 more figures