SAMBAR is a algorithm for Vision-Language-Action (VLA) models that prevents when learning new tasks sequentially. It works by casting continual learning as a constrained optimization problem solved with the method of multipliers, using a dual variable to penalize parameter drift from previous values. Additionally, it selectively anchors parameters critical to prior tasks while leaving others free for new learning. The method is evaluated on the LIBERO simulation benchmark and on hardware, showing that it retains all learned tasks while replay-free baselines completely forget the first task. This is interesting because deployed VLA policies need to continually acquire new skills without storing old task demonstrations.
Replay-free continual learning method prevents forgetting in vision-language-action models
SAMBAR combines the method of multipliers with selective parameter anchoring to let VLA policies learn new tasks while preserving old ones.
Top University
Aayushi Shrivastava · Xunlan Zhou · Hongrui Zhao · Ziyu Chen · Negar Mehr
University of California Berkeley
Research Digest··3 min read
Thread:Robot World Models
In one line
SAMBAR prevents catastrophic forgetting in VLA models during continual learning without needing prior task data.
Before you read this
Three ideas this paper builds oncontinual learningA learning paradigm where a model sequentially acquires knowledge from multiple tasks without significantly forgetting earlier ones.72 papers →vision-language-action modelA robot policy model that maps visual observations and language instructions to actions, trained on demonstrations and often used for manipulation tasks.87 papers →catastrophic forgettingLoss of previously learned knowledge when a model is fine-tuned on new tasks.7 papers →
Read it as
The authors introduce SAMBAR, a continual learning algorithm for vision-language-action (VLA) models that requires no replay of previously seen demonstrations.
The authors address in VLA policies, which commonly occurs when a pretrained model is finetuned sequentially on new manipulation tasks.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.