DRAP paper proposes replay-free continual learning that cuts plasticity at dynamics-risk hotspots for physical AI
Oswaldo Ludwig · hf · 2026-11-05
A new paper by Oswaldo Ludwig introduces DRAP (Dynamics-Risk Adaptive Plasticity), a replay-free continual learning method for physical AI. It estimates the historical importance of adaptive tensors from sample-level gradients weighted by local trajectory-amplification risk, using a ridge-regularized local linear model whose largest singular value serves as a Lyapunov-inspired, policy-independent risk estimate of one-step perturbation amplification.
After each task, risk-weighted squared gradients accumulate in a tensor-level importance memory, which scales each tensor's effective AdamW learning rate to reduce plasticity of components supporting past behaviors at dynamically sensitive trajectory points. Unlike checkpoint-based consolidation, protection is active from the first update of each new task. The paper instantiates DRAP on a fixed KappaTune-selected subnetwork of OpenVLA, improving retention of past behaviors without sacrificing new-task learning.
More from Embodied
- Chinese AI Scientist completes 11,382-step experiment in 203 hours autonomously — yogthos · 2026-11-05
- Beff Jezos shares IRL photos of Tesla Cybertruck at NASA Florida event — beffjezos · 2026-11-05
- HandsOn benchmark: best dexterous manipulation policy manages only 29.1% success on Core-50 — jiqizhixin · 2026-11-05
- Aina's smart ring aims to be a one-tap remote control for your AI agents — Wired AI · 2026-11-05
- Waymo to end Uber exclusivity in Austin and Atlanta, launch own app in January 2028 — emmanuelvivier · 2026-11-05
- Tired of screens and headsets, this maker built a wearable computer hat — aaraalto · 2026-11-05