The one-paragraph version
This paper asks where a language-model robot policy gets its ability to get better at physical tasks by trying, failing, and rewriting itself. The authors build PhysRSI-eval, a test where six vision-language-model systems plus a partial Gemini 4 run each attempt 18 dual-arm manipulation tasks in simulation from the RoboDojo suite, with model weights frozen but a persistent workspace of code, tools, and notes that carries over across tries. Each task allows one first attempt without that workspace, called K0, followed by up to 17 refinement rounds, called K1 through K17, stopping at the first native success. They score every system on five separate axes rather than one final score: how good the first try is, how early gains appear, what score is finally attained, how much physical budget is saved, and how many early failures are later fixed. From exploratory rank comparisons they propose the Physical Barrel Hypothesis: like staves of a barrel, different underlying capabilities limit different axes, with spatial reasoning most tied to starting well and perception most tied to finishing well and recovering. They then show two kinds of observational prediction are possible in this small sample — public benchmark profiles predict the ordering of final scores and gain timing, and an improvement-curve shape learned from one model helps forecast another model from only its first five rounds — and they show retained experience helps beyond just retrying, with frozen workspaces still working on unseen layouts and sometimes speeding up a different task.
The problem they're solving, and why it matters
Frontier agents can now write robot code directly, watch it fail, and try to fix it. The authors call this physical recursive self-improvement, meaning physical RSI for short: improvement without weight updates, only by changing the workspace the agent leaves for its future self. The field already has demos of this working, and even a reported failure where a self-written fix worked once but broke on repeat. What is missing, they argue, is measurement of the process itself. A single endpoint score hides when improvement happened, what it cost in physical tries, whether retrying alone would have done as well, and whether what was learned was a reusable trick or just overfitting to one scene. Without separating starting skill from learning from feedback, we cannot say what capability to improve next, how much interaction budget a deployment will need, or whether to trust a learned routine elsewhere. This matters practically because physical tries are expensive, slow, and risky compared to chatting, and scientifically because self-improvement is often talked about as one thing when it may be several different bottlenecks.
The key idea, in plain words
The key idea is to split physical self-improvement into five measurable outcomes, then ask which outside capabilities line up with which outcome. Think of first trying to pack objects into a box. Doing okay immediately might depend on judging where things are in space and how the gripper should be oriented. Coming back after dropping something and figuring out what went wrong might depend more on seeing and interpreting the failure image and grader message. Being good at writing code in general might be necessary to implement any fix, but not enough to know what to fix. That is the intuition behind the Physical Barrel Hypothesis: a system is like a barrel made of staves for spatial reasoning, perception, coding, and harness operation, and the shortest stave depends on which part of improvement you measure. Starting well and learning from failure may be limited by different staves, so two systems with similar average public benchmark scores can still behave very differently when given the same physical budget. The second key idea is observational scaling: instead of claiming a universal law from training compute, which they do not have for closed systems, they test whether you can predict the order of systems and the shape of learning curves from public scores plus a little early physical evidence.
How it actually works, step by step
First, they fix the task set and difficulty spread. They take 18 RoboDojo tasks covering arranging, stacking, placing in containers, insertion, pouring, pushing, and conveyor picking, such as Cover Blocks, Deposit Coin, Insert Key, and Organize Table. They rank tasks by official RoboDojo success rates and pick three per stratum from easiest, called D1 with about 60 to 68.7 percent official success, to hardest, called D6 with 0 to 0.4 percent, so the suite has both immediate competence and headroom.
Second, they fix the interaction loop. In the first-attempt stage, the agent gets one in-context episode of looking, reasoning, and acting through shared robot tools, with no persistent workspace. Scores here are averaged over three runs per task. In the self-improvement stage, each round starts from the committed workspace of skills, tools, policy code, and memory, runs one judged physical attempt that returns a native score between 0 and 1 plus a success flag and failure reason, then revises the workspace with robot actions turned off and commits it for the next round. The change from K0 to K1 also adds workspace tools, so the authors treat K1 to K17 as the clean self-improvement segment.
Third, they compute the five axes after averaging task curves. In words first, then symbols: First-Attempt Performance is simply mean K0 score, called A1 equals q0. Efficiency is how early sustained gains appear relative to total gain, called A2, which averages how far above start each middle round is divided by final gain. Success is the average score once the system first reaches 90 percent of its peak gain, called A3, not a binary rate and not an assumed ceiling. Budget is physical savings on tasks it actually solved, called A4, based on canonical steps used plus penalties for timeouts and extra attempts relative to the per-task step limit and 17-round budget. Improvement is the share of tasks failed at K1 that are later solved, called A5 equals CH minus C1 over one minus C1, where Ck is fraction solved by round k.
Fourth, they run capability diagnosis, scaling prediction, and generalization tests. For diagnosis they group 17 public scores collected in October 2026 into Spatial, Harness, Perception, and Coding and compare ranks with A1 through A5. For scaling they try to predict held-out order of best score by K17 and order of half-gain timing, meaning the first round where best-so-far reaches half its K1 to K17 gain, called T50. They also fit a Weibull growth curve shape from Astra and calibrate only its start, plateau, and relative speed on another model’s first few rounds to forecast the rest. For generalization they ablate what history the agent sees, freeze K6 workspaces and test them on five unseen layouts with no further revision, and seed a new task with workspaces pooled from other tasks.
What they showed — results, honestly, with limitations
The five axes diverge in informative ways. GPT-6 Astra led on all five and solved 15 of 18, followed by GPT-6.1 Sol with 12, Claude Opus 5.5 with 13, Claude Sonnet 5.5 with 11, GPT-6 Sol with 7, and GPT-5.6 Sol with 1. Gemini 4 in a separate ten-task run solved 6. Crucially, a higher final score did not mean earlier or cheaper improvement: Opus beat GPT-6.1 Sol on Success, 0.744 versus 0.703, but lost on Efficiency, 0.609 versus 0.822, and on Budget on solved tasks, 0.579 versus 0.808. Task-mean scores rose from 0.01 to 0.37 at K0 to 0.10 to 0.84 at the end, but GPT-5.6 Sol stayed near zero success for 16 rounds, showing more tries do not close every gap in this budget.
The capability associations that motivate the barrel story are strong in rank terms but explicitly exploratory and not causal. Spatial reasoning had the strongest observed rank association with First-Attempt Performance at rho equals 0.94 and with Efficiency and Budget at 0.83, while Perception led for Success and Improvement at 0.94, and Coding was weakest everywhere at 0.37 or less. The paper is clear, and the appendix confirms, that with only six systems and exact permutation tests, the smallest adjusted q-value is 0.11, so this does not pass conventional significance after correction. It is a candidate-bottleneck hypothesis, illustrated by GPT-5.6 Sol beating Astra on LiveBench coding yet solving only one task.
The scaling demonstrations are also small-sample and observational. Ranking by equal-weight capability profiles reproduced the observed order of best score by K17, and a kernel regressor trained on four systems recovered all pairwise orders of who reaches T50 first. For forecasting, calibrating Astra’s fitted shape on GPT-6.1 Sol’s K1 to K5 forecast K6 to K17 with mean absolute error 0.019 versus 0.112 for just holding the last score. That is impressive as a proof of concept, but it is one transfer pair on 18 tasks, not a validated law.
On overfitting versus generalization, the evidence supports a mixed answer: both scene-specific adaptation and reusable experience. In a six-task Astra ablation with changing layouts every round, late-round K4 to K6 success was 15 of 36 with full history versus 3 of 36 with no history, and removing grader feedback hurt most, down to 5 of 36. Frozen K6 workspaces with no further revision succeeded in 25 of 60 trials on unseen layouts using files alone, and 23 of 60 when the test agent could also read old raw logs, suggesting no added benefit from raw history at test time in this subset. Three cross-task cases reached first success earlier when seeded from other tasks, for example Opus on Organize Table from K14 to K7 and on Deposit Coin from K15 to K8, and Sonnet on Cover Blocks from K9 to K1. Tool traces add texture: Opus, Sonnet, and GPT-6 Sol built large tool libraries with different breadth of reuse, while Astra and GPT-6.1 Sol generated about 99 percent of actions directly and relied more on notes and memory. Limits the authors state plainly include one chain per task in the main comparison, no variance over full chains, first-success stopping rather than reliable repetition, conditional Budget and Improvement sets that differ by model, and transfer tests limited to Astra and a few tasks and cases.
Why this might matter to me
For spatial reasoning in LLMs and VLMs, the split between starting well and recovering from failure is directly useful. If spatial scores track first-attempt and early-gain axes while perception tracks recovery, then evaluating a VLM only on single-shot question answering misses the perception-for-debugging skill that matters after a robot fails. It suggests adding benchmarks where the model must localize what went wrong from images and relate that error vector to the next action, not just label a scene.
For world models, the forecasting result is a hint that early physical trajectories carry predictive shape. A compact growth template plus start, plateau, and speed calibrated from a few rounds predicted much of the rest. That is conceptually close to learning a task-level world model of progress, where physical feedback updates beliefs about how fast a policy will converge and when to stop spending budget.
For text-to-CAD and parametric geometry, the tool-library findings map neatly. Opus accumulated a broader reusable toolbox while others reused a narrow subset, and one transfer example combined a parameterized move tool from packing with a memorized pitched grasp note to extend reach in organizing. That is analogous to parametric CAD operations plus operating knowledge about when parameters work. It suggests representing learned skills as typed, parameterized procedures with preconditions learned from failure, rather than opaque scripts.
For 3D generation and scene representation, the frozen-workspace result matters: files alone stayed useful on unseen layouts about 42 percent of the time, but performance dropped from about 88 percent on the training layout in the reported comparison. Scene-specific tuning and transferable structure coexist, which argues for scene representations that separate reusable object-level routines from layout-specific placements, so adaptation can keep the former and refit only the latter.
Terms worth knowing
Physical RSI means physical recursive self-improvement: the agent improves by revising its own persistent code, tools, and notes from physical trial and error, with model weights frozen.
PhysRSI-eval is the authors’ protocol to measure that: K0 first attempts with no persistent workspace, then Harness rounds K1 to K17 with one judged physical attempt plus offline revision per round, stopping at first native success.
Five-axis profile means the five outcomes A1 through A5: First-Attempt Performance, Efficiency or gain timing, Success or attained high-window score, Budget or physical savings on solved tasks, and Improvement or share of K1 failures later solved.
Physical Barrel Hypothesis is the proposal that different axes have different capability bottlenecks, so strong coding cannot always compensate for weak spatial reasoning or perception.
Observational scaling here means predicting physical improvement from public benchmark profiles and early physical curves without training-compute data, as opposed to a causal or compute-based scaling law.
T50 and tau-90 are timing markers: T50 is the half-gain round when best-so-far first reaches half its K1 to K17 gain, while tau-90 is when the mean first reaches 90 percent of peak gain and defines the start of the Success window.
Workspace, ablation, and frozen transfer refer to what the agent retains and how it is tested: the workspace is the committed files carried between rounds, ablation removes channels like grader feedback or images to see what helped, and frozen transfer reuses a workspace on new layouts or tasks with revision turned off to test generalization.