Researchers led by Jungjin Lee have presented Physics-Informed Causal Reasoning (PICR), a model that lets a robot adapt to changed physical conditions from its own experience, without retraining or a hand-written model of the robot, according to a paper on arXiv. Systems that look the same in a snapshot can react differently to the same action because of conditions the snapshot hides, the authors write.

PICR re-estimates only a context from recent interactions, and that context steers a fixed physics-informed network that works out the robot's inertia and potential energy. On a Franka Panda arm simulated in NVIDIA Isaac Sim, it predicted motion under changing and unseen payloads, cutting error on unseen ones by 1.6 to 3.8 times compared with baselines, and it could answer what-if questions about other torques.

Source: arXiv Robotics (cs.RO)

Background

Jungjin Lee

Robotics researcher

Robotics researcher.