Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a method for rapidly adapting neural robot dynamics models by combining offline training with efficient online updates. An incremental neural dynamics model is learned offline, and then a low-rank second-order parameter adaptation is performed online, allowing for quick adjustments without full retraining. The approach is validated on a real quadrotor, demonstrating robust predictive tracking control in new operational conditions.
Quadrotors can now adapt their dynamics models on-the-fly, enabling more robust tracking control in changing real-world conditions.
Accurate dynamics models are critical for the design of predictive controller for autonomous mobile robots. Physics-based models are often too simple to capture relevant real-world effects, while data-driven models are data-intensive and slow to train. We introduce an approach for fast adaptation of neural robot dynamic models that combines offline training with efficient online updates. Our approach learns an incremental neural dynamics model offline and performs low-rank second-order parameter adaptation online, enabling rapid updates without full retraining. We demonstrate the approach on a real quadrotor robot, achieving robust predictive tracking control in novel operational conditions.