Paper studies physical recursive self-improvement in VLM policies
Original title: Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
PhysRSI-eval scores VLMs on 18 manipulation tasks under a fixed budget and tests whether public scores predict later physical gains.
The paper investigates mechanisms, scaling, and generalization of physical recursive self-improvement. PhysRSI-eval uses five axes: first-attempt performance, efficiency, success, budget, and improvement from failure. Different VLMs are tested on 18 manipulation tasks with a fixed interaction budget. Workspace interventions show that removing history cuts late-round success, frozen workspaces stay useful on unseen layouts, and three cross-task cases succeed earlier.
Unconfirmed: The Physical Barrel Hypothesis and coexistence of adaptation with reusable experience are interpretive claims from the experiments.
- The Physical Barrel Hypothesis says different capabilities constrain different parts of improvement.
- High public benchmark averages need not imply early or efficient physical gains.
- The work predicts relative VLM-agent performance from public capabilities and later progress from early feedback.
- Scene-specific adaptation and reusable experience can coexist, the authors suggest.
From the original article
1 min read at arxiv.orgIn 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 · arxiv.org
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.