Paper studies physical recursive self-improvement in VLM agents
Original title: Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
PhysRSI-eval scores 18 manipulation tasks and finds close public averages can hide late or inefficient gains.
The paper examines mechanisms, scaling, and generalization of physical recursive self-improvement for vision-language models used as policies. PhysRSI-eval uses five axes, and different VLMs are tested on 18 manipulation tasks under a fixed interaction budget. Workspace interventions suggest history matters for late success, while frozen workspaces remain useful on unseen layouts.
Unconfirmed: The Physical Barrel Hypothesis is the authors' interpretation of their results.
- Axes cover first attempt, efficiency, success, budget, and improvement from failure.
- Authors propose a Physical Barrel Hypothesis on capability bottlenecks.
- Early physical feedback is used to forecast later progress.
- Removing history reduced late-round success.
- Three cross-task cases reached success earlier.
From the original article
1 min read at huggingface.co
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.
By Zekai Li, Jiayi Yu, Jinlu Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.