This paper addresses deformable object manipulation challenges in robotics by introducing Physics-Aware Domain Randomization (PADR), which varies deformable material properties based on physically plausible distributions. PADR integrates tactile feedback with visual observations and evaluates on cloth folding and object manipulation tasks, demonstrating a 91% success rate in sim-to-real transfer.
Key findings
PADR outperforms uniform domain randomization by 20 percentage points and no domain randomization by 39 percentage points in sim-to-real transfer.
The approach reduces sensitivity to material stiffness variations by 82% compared to baseline methods.
Physics-aware sampling and tactile feedback significantly contribute to performance.
The method enables robust zero-shot transfer to unseen real-world objects.
Limitations & open questions
The study focuses on specific material categories and may not generalize to all types of deformable objects.
The effectiveness of PADR in diverse real-world scenarios with varying environmental conditions is not fully explored.