DRAP: A Replay-Free Continual Learning Method
What happened
The study proposes DRAP, a replay-free continual learning method: it uses locally trajectory-amplified risk-weighted action gradients to estimate the historical importance of model tensors, and accordingly scales the effective AdamW learning rate of each tensor. The method is instantiated on a fixed subnetwork of OpenVLA selected via KappaTune, and performs sequential adaptation on the Spatial, Object, and Goal suites of LIBERO. The authors claim that, compared with standard AdamW fine-tuning and EWC, DRAP significantly improves retention of old behaviors while maintaining the ability to learn new suites.
AI 根据报道生成 · 6 days ago更新
Reporting timeline
Follow the reporting to explore different perspectives.
- HuggingFace Daily PapersDRAP: 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.
Activity for this event
- Current
- 3
- Peak
- 3Nov 5 21:00
- Past 24 hours
- –
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.