Where is Physical RSI capability from?An Observational Scaling Law of LLM-as-policy

Overview

Benchmark

A controlled loop of physical attempts

18

RoboDojo manipulation tasks

K0 + 17

first attempt + self-improvement rounds

5

axes of physical RSI

6 + 1

systems (+ Gemini 4 on ten tasks)

• A1First-Attempt Performance
• A2Efficiency
• A3Success
• A4Budget
• A5Improvement

Benchmark

Tasks across difficulty levels

Q1: Where is Physical RSI capability from?

Physical Barrel Hypothesis

Different aspects of improvement may have different capability bottlenecks: spatial reasoning for starting well, perception for recovering from failure. Strong coding cannot always compensate for a weakness elsewhere.

Q1: Where is Physical RSI capability from?

Five-axis profiles

Q1: Where is Physical RSI capability from?

Five-axis profiles, as drawn in the paper

Q2: How can we quantify Physical RSI and its scaling law?

Higher attained scores can come with later gains and greater physical costs: Opus 5.5 beats GPT-6.1 Sol on Success (0.744 vs 0.703) but trails on Efficiency (0.609 vs 0.822) and Budget (0.579 vs 0.808).

Five-axis profiles. K0 scores are averaged over three runs per task. *Gemini 4 is evaluated on ten tasks with one K0 run per task.

Q2: How can we quantify Physical RSI and its scaling law?

Public capability and physical RSI separate differently

Across the four GPT systems, mean public benchmark scores span 24 points; the physical profile mean spans 59.5 points. GPT-5.6 Sol stays near zero cumulative success for 16 rounds.

Q2: How can we quantify Physical RSI and its scaling law?

How far, and how early?

Q2: How can we quantify Physical RSI and its scaling law?

Predicting the order from public capability profiles

Classical scaling laws relate performance to model size and training compute, which closed-source models do not disclose. Observational scaling instead uses public benchmark scores, aggregated into spatial reasoning, perception, coding and harness-use factors, as an observable substitute. Public capability profiles reproduce the observed order of both the best score by K17 and T₅₀, the half-gain round.

Q2: How can we quantify Physical RSI and its scaling law?

What can early physical feedback predict?

A Weibull template fitted to Astra, calibrated on GPT-6.1 Sol's K1–K5, forecasts K6–K17 with mean absolute error 0.019, versus 0.112 for holding the last observed score.

Q3: Does Physical RSI reflect overfitting or generalization?

Q3: Does Physical RSI reflect overfitting or generalization?

Does retained experience help beyond repeated attempts?

Astra ablation on six tasks, two chains each, layouts changing every round. Late-round success: 41.7% with the full loop, 8.3% without history.

Per-round online success in the Astra ablation. Successes out of 12 chains per condition and round; K1–K6 run in full.

Q3: Does Physical RSI reflect overfitting or generalization?

What survives when revision stops?

Astra workspaces frozen after K6 and run on five unseen layouts, with no revision or feedback.

Frozen held-out test of Astra K6 workspaces. 12 sources × 5 held-out layouts for the complete conditions; fixed-layout sources reach 45.0% on held-out layouts versus 88% (14/16) on their own layout.

Q3: Does Physical RSI reflect overfitting or generalization?

Can experience help on a different task?

Sonnet 5.5 and Opus 5.5 encode control in self-written tools that generate over 96% of later-round actions; Opus reuses the broadest toolbox (lineage reuse 0.347 vs 0.167 for Sonnet and 0.086 for GPT-6 Sol).

Citation

@misc{li2026physicalrsicapabilityfrom,
  title={Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy},
  author={Zekai Li and Jiayi Yu and Jinlu Zhang and Luyang Niu and Zekun Qi and Jun Zheng and Xuchuan Chen and Dairu Liu and Yintianrun Zhang and He Wang and Li Yi},
  year={2026},
  eprint={2610.26848},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.26848}
}