DRAP: Dynamics-Risk Adaptive Plasticity for Continual Learning in Physical AI

The study proposes DRAP, a replay-free continual learning method that uses locally trajectory-amplified, risk-weighted action gradients to estimate the historical importance of model tensors and scales each tensor's effective AdamW learning rate accordingly. The method is instantiated on OpenVLA's KappaTune subnetwork and sequentially adapted on the LIBERO Spatial, Object, and Goal suites; compared with standard AdamW fine-tuning and EWC, it markedly improves retention of old behaviors while maintaining the ability to learn new suites.