Nvidia, KAIST's ACCRUE Cuts Real-Robot RL Training 31-68.5%
TL;DR
- ACCRUE carries lightweight actor and critic networks across tasks so each new skill reuses behavioral and value knowledge from prior RL stages.
- On contact-rich assembly in simulation and on real hardware, the framework cuts real-robot training time by 31.0% to 68.5%.
- Prior behavior is preserved by distilling the actor toward cached targets on a small number of retained episodes, with near-zero forgetting.
A new paper from NVIDIA and KAIST reports that a continual off-policy reinforcement learning framework can cut the time a real robot needs to pick up a new assembly skill by between 31.0% and 68.5%.
Posted to arXiv on October 13 by Changyeon Kim (NVIDIA, KAIST), Kimin Lee (KAIST), and NVIDIA's Yashraj Narang and Yijie Guo, ACCRUE "carries lightweight actor and critic networks across tasks, allowing each stage to reuse behavioral and value knowledge acquired through prior RL," the authors write. The inherited critic adapts only on current-task transitions and guides current-task actor learning, while prior behavior is preserved "by distilling the actor toward cached targets on a small number of retained episodes."
The headline claim covers contact-rich assembly sequences in simulation and on real-robot hardware, with "near-zero forgetting." The abstract publishes no per-task breakdown inside the 31.0% to 68.5% band, no base-policy name, and no parameter count for the inherited actor and critic.
ACCRUE lands in a crowded week for continual-RL and manipulation work on our tracker: Sber's GreenWAM and NVIDIA's own TacitVLA both landed within a day.
Originally reported by arxiv.org
Read the original article →Original headline: KAIST and Nvidia's ACCRUE Cuts Real-Robot Training Time 31-68% Via Continual Off-Policy RL