DRAP cuts plasticity in OpenVLA for replay-free continual learning
Original title: DRAP: Dynamics-Risk Adaptive Plasticity for Continual Learning in Physical AI
Risk-weighted gradient memory scales each tensor’s AdamW rate and retains prior LIBERO behaviors better than EWC.
DRAP is a replay-free continual-learning method for physical AI that estimates tensor importance from sample-level action gradients weighted by local trajectory-amplification risk. A ridge-regularized local linear model of expert state-action transitions supplies a policy-independent, Lyapunov-inspired one-step amplification estimate. After each task, risk-weighted squared gradients accumulate in a tensor-level memory that later scales AdamW learning rates. On a fixed KappaTune-selected subnetwork of OpenVLA, sequential adaptation across LIBERO Spatial, Object, and Goal retains prior behaviors better than standard AdamW and EWC while still acquiring new suites.
- Small per-step policy errors can accumulate and cause closed-loop task failure.
- Importance uses the largest singular value of a fitted state-transition matrix.
- Plasticity drops for tensors that mattered at dynamically sensitive trajectory points.
- Evaluation is replay-free sequential adaptation on three LIBERO suites.
From the original article
1 min read at huggingface.co
Continual learning in physical AI differs from conventional continual learning because small per-step policy errors can accumulate over a trajectory. Each deviation may move the agent outside the training distribution, increase subsequent prediction errors, and ultimately cause closed-loop task failure.
By Oswaldo Ludwig · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.