Physical recursive self improvement shows varied learning limits in robots
Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
RoboticsArtificial IntelligenceComputer Vision and Pattern Recognition
Summary
This paper looks at how robots and virtual agents get better at physical tasks by learning and improving themselves. The authors created a way to measure this improvement using five different criteria, like how quickly and efficiently a robot learns. They found that different skills limit improvement in different ways, and early success doesn't always mean fast learning. The study also shows that robots can reuse what they've learned in different settings and that environments can impact their progress.
What this means in practice
- •For robotics engineers: Design metrics and testing protocols for evaluating how robots self-improve on physical manipulation tasks under fixed resource limits.
- •For industrial automation teams: Optimize robot training procedures by tracking early performance to forecast later improvements in task success across changing workspaces.
Authors
Zekai Li, Jiayi Yu, Jinlu Zhang, Luyang Niu, Zekun Qi, Jun Zheng, Xuchuan Chen, Dairu Liu, Yintianrun Zhang, He Wang, Li Yi
Abstract
In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.