Measuring the path to physical self-improvement
Recursive self-improvement in physical agents remains poorly understood. While language models have well-charted scaling laws, their robotic counterparts lack comparable observational frameworks. This paper systematically investigates how physical RSI emerges, scales, and generalizes across manipulation tasks, introducing PhysRSI-eval—a five-axis metric that captures First-Attempt Performance, Efficiency, Success, Budget, and Improvement from failure.
Five-axis quantification of physical RSI
The proposed metric moves beyond simple success/failure reporting. First-Attempt Performance measures how well an agent solves a task without prior trial-and-error. Efficiency captures the relationship between interaction budget and final success. Success reflects the ultimate task completion rate. Budget tracks the resources consumed to achieve a given performance level. Improvement from failure quantifies how much later attempts benefit from early mistakes. Together, these axes can distinguish systems with nearly identical public benchmark averages, revealing that high overall scores may mask poor early learning or wasteful resource use.
The Physical Barrel Hypothesis
Analysis of experimental results—evaluating different VLMs on 18 manipulation tasks under a fixed interaction budget—motivates the Physical Barrel Hypothesis. This hypothesis proposes that different capabilities (such as perception, affordance understanding, or motor planning) may constrain different parts of the improvement curve. Some capabilities may enable rapid early gains, while others contribute more to late-stage polish. The five-axis metrics make this distinction visible: a system might score well on overall success but poorly on Efficiency or Improvement from failure, indicating that its strengths lie in a narrow band of the improvement curve.
Observational scaling from public capabilities to physical progress
The study explores two complementary scaling perspectives. First, it asks whether public VLM capabilities (as measured on standard vision-language benchmarks) can predict the relative performance of VLM agents on physical manipulation tasks. The results suggest only moderate predictability: a model strong in abstract reasoning does not necessarily translate to strong physical RSI, underscoring the gap between semantic knowledge and grounded interaction.
Second, the paper forecasts later progress from early physical feedback. By analyzing how early interaction data predicts eventual success, the authors identify conditions under which agents improve more quickly. Three cross-task cases achieve success earlier than expected, suggesting that some skills transfer across manipulation domains even when the specific skills do not.
Workspace interventions and the residue of improvement
Beyond scaling, the authors conduct workspace interventions to understand what improvement leaves behind. Removing interaction history reduces late-round success, indicating that accumulated experience matters for finishing tasks. Frozen workspaces—environments preserved after training—remain useful on unseen layouts, pointing to the value of reusable experience rather than purely scene-specific adaptation. These findings suggest that scene-specific adaptation and reusable experience can coexist, with each contributing to different aspects of physical RSI.
Implications for robotics research
For researchers building LLM-as-policy systems, the PhysRSI-eval framework provides a more nuanced view of agent progress than success rate alone. The Physical Barrel Hypothesis offers a diagnostic lens: when an agent plateaus, the five-axis metrics can reveal whether the bottleneck is early learning, inefficient resource use, or poor improvement from failure. The coexistence of scene-specific and reusable signals also informs how future systems should balance memorization of specific contexts versus learning generalizable skills.
While the evaluation is currently limited to 18 manipulation tasks and a fixed set of VLMs, the introduced metrics and hypotheses are broadly applicable to any physical agent trained via interaction. The paper opens the door to more targeted interventions—such as selectively preserving history or freezing specific workspace regions—depending on which axis of improvement a researcher wishes to enhance.