Universal Controllers with Differentiable Physics for Online System Identification

Michelle Guo, Wenhao Yu, Daniel Ho, Jiajun Wu, Yunfei Bai, Karen Liu, Wenlong Lu

Abstract:

Creating robots that can handle changing or unknown environments is a critical step towards real-world robot applications. Existing methods tackle this problem by training controllers robust to large ranges of environment parameters (Domain Randomization), or by combining ``Universal'' Controllers (UC) conditioned on environment parameters with learned identification modules that (implicitly or explicitly) identify the environment parameters from sensory inputs (Domain Adaptation). However, these methods can lead to over-conservative behaviors or poor generalization outside the training distribution. In this work, we present a domain adaptation approach that improves generalization of the identification module by leveraging prior knowledge in physics. Our proposed algorithm, UC-DiffOSI, combines a UC trained on a wide range of environments with an Online System Identification module based on a differentiable physics engine (DiffOSI). We evaluate UC-DiffOSI on articulated rigid body control tasks, including a wiping task that requires contact-rich environment interaction.
Compared to previous works, UC-DiffOSI outperforms domain randomization baselines and is more robust than domain adaptation methods that rely on learned identification models. In addition, we perform two studies showing that UC-DiffOSI operates well in environments with changing or unknown dynamics. These studies test sudden changes in the robot's mass and inertia, and they evaluate in an environment (PyBullet) whose dynamics differs from training (NimblePhysics).