ACCRUE: Continual Off-Policy RL for Robot Foundation Models
Loading video
Loading videoACCRUE keeps a robot foundation model frozen and trains only a compact actor-critic with off-policy RL from real robot interaction, inheriting the actor-critic across tasks and distilling earlier behavior so experience accumulates without erasing prior skills. On the AutoMate and NIST manipulation suites, later tasks reach 95 percent success in 4-12 minutes of robot interaction, cutting interaction by 31.0-68.5 percent versus single-task RL while prior tasks end at 98.3 percent success.
Reinforcement LearningContinual learningRobot foundation modelsManipulationOff-Policy RLAutoMateNIST
Category: arm
Author: @cykim1006
Date: 2026-10-14T00:00:00
Duration: 59.999s
Reference: https://arxiv.org/abs/2610.18596
