UniIntervene++: An Adaptive Intervention Agent for Efficient Real-World Reinforcement Learning
The paper reports a robot-learning intervention system that cuts human control to 0.77% while reaching 89.67% average success across five manipulation tasks.
UniIntervene++ treats autonomous policy execution, trajectory correction, and a task-structured CodePolicy as options whose values are learned online. It periodically lets the RL policy run without assistance to test whether intervention is still needed. Assisted behavior also feeds training data back into the RL policy, changing later intervention choices as the policy improves. ArXiv · AI/CL/LG's note
UniIntervene++ treats autonomous policy execution, trajectory correction, and a task-structured CodePolicy as options whose values are learned online. It periodically lets the RL policy run without assistance to test whether intervention is still needed. Assisted behavior also feeds training data back into the RL policy, changing later intervention choices as the policy improves. ArXiv · AI/CL/LG's note
score 4