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. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.
Figures & tables
Figure 1: PhysRSI-eval at a glance. (Q1) Capability sources and the Physical Barrel Hypothesis. (Q2) Five-axis quantification and observational scaling of performance, gain timing and model rankings. T50 is the half-gain round, the first round at which the best mean score reaches half of its K1–K17 gain. (Q3) Overfitting and generalization across tasks and scenes.
Figure 2: Two stages, one interface. Top: the task input and robot interface shared by both stages (observations, physical actions, memory and files, history). Left, Evaluate (first-attempt round K0): one in-context episode of observe–reason–act through the shared tools, with no workspace; history and notes stay inside the episode. Right, Optimize (self-improvement rounds K1–K H ): each round runs from a persistent workspace of skills, tools, policy and memory, is graded natively, revised with robot actions closed, and committed as the start of the next round, until the first native success or the round limit.
Figure 3: Difficulty-balanced task sampling. Three tasks per difficulty stratum, from D1 (highest official RoboDojo success rate) to D6 (lowest), covering RoboDojo’s five official task categories. Strata are assigned by ranking the tasks on the official RoboDojo success rate [ 3 ] and are fixed before the main evaluation; they do not describe the subsequent refinement outcomes.
Figure 4: Capability associations and the barrel hypothesis. The matrix reports exploratory rank associations across six systems on the 18 tasks. Each barrel represents one RSI axis, with a stave per capability group. Stave heights illustrate a proposed bottleneck interpretation of the associations.
System
Tasks
A1
A2
A3
A4
A5
GPT-6 Astra
18
0.3676
0.8594
0.8509
0.8429
0.7500
GPT-6.1 Sol
18
0.2843
0.8224
0.7028
0.8081
0.5714
GPT-6 Sol
18
0.0306
0.4658
0.4722
0.3718
0.3889
GPT-5.6 Sol
18
0.0120
0.2247
0.1037
0.3013
0.0556
Claude Opus 5.5
18
0.1769
0.6091
0.7444
0.5794
0.6875
Claude Sonnet 5.5
18
0.1241
0.6946
0.7006
0.6984
0.5625
Table 1: Five-axis profiles. K0 scores are averaged over three runs per task. A1–A5 denote First-Attempt Performance, Efficiency, Success, Budget and Improvement. ∗ Gemini 4 is evaluated on ten tasks with one K0 run per task.
Figure 5: Five-axis profiles. Left: seven overlaid profiles. Right: one panel per system. Axes clockwise from the top are First-Attempt Performance (A1), Budget (A4), Success (A3), Efficiency (A2) and Improvement (A5), on a shared 0–1 scale. The six reference systems use all 18 tasks, as in Table 1 ; ∗ Gemini 4 uses its ten evaluated tasks ( Section 3.3 ).
Figure 6: Public capability and physical RSI separate differently. Left: cumulative success during refinement on 18 tasks; the shaded region marks the gap between GPT-6.1 Sol and GPT-6 Sol. Right: mean public score over 17 selected benchmarks versus 100×(A1+A2+A3+A4+A5)/5 . The physical summary is an arithmetic mean, not radar area.
Figure 7: Predicting the timing of gains from public capability profiles. Left: four capability profiles of each system, computed from 17 public benchmarks as normalized ranks among the six systems. Right: best mean score qˉk on the 18 tasks. Circles mark T50 , the first round at which qˉk reaches half of its K1–K17 gain. Labels give predicted / observed ranks: T50 beside the circles and best score reached by K17 on the right. Each pair of systems is predicted from the other four.
Figure 8: Forecasting GPT-6.1 Sol from Astra’s improvement shape. Left: Astra’s mean score and the Weibull template fitted to it. Right: GPT-6.1 Sol with K1–K5 available. The template is calibrated on these rounds and forecasts the hidden rounds through K17; shading shows the 90% paired-task bootstrap band. All 18 tasks are used.
Figure 9: Cross-task workspace transfer. Top: original self-improvement trajectories for tasks first solved at K6 or later. Bottom: trajectories initialized from a workspace pooled from all other tasks, excluding the target task. First success advances from K9 to K1 for Sonnet 5.5 on Cover Blocks, from K14 to K7 for Opus 5.5 on Organize Table, and from K15 to K8 for Opus 5.5 on Deposit Coin. These examples illustrate faster adaptation with transferred workspaces. Red crosses indicate unsuccessful attempts, dark checks mark first success, and pale checks carry success forward without additional evaluation.
Figure 10: Tool Libraries and Reuse across Models. (a) Total tool counts and the proportions of perception and geometry tools versus action tools. (b) Fraction of tool lineages reused in subsequent rounds, among those with later rounds available. Reuse includes direct use, modification and derivation of existing tools. (c) Reuse breadth, defined as one minus the share of subsequent calls accounted for by the top 10% of reused lineages. Higher values indicate reuse distributed across more lineages. Astra is omitted because it creates very few tools, used only for verification. GPT-6.1 Sol’s estimates reflect a small library of seven tools.
#
Task
Steps
Scoring (tiers, %)
1
arrange_largest_number
1050
graded [5,15,30,100] or [5,15,25,40,100]
2
classify_objects_by_language
1100
graded [10,40,100]
3
cover_blocks
800
graded [0,0,5,15,30,100]
4
deposit_coin
300
graded [20,100]
5
fill_pen_holder
1100
graded [10,25,40,90]
6
insert_key
300
graded [15,100]
Table 2: The 18 PhysRSI-eval tasks. Step limits are per round (25 Hz). Graded tasks score the highest tier reached (tiers in percent of the native score); binary tasks score 0 or 1. The success check is independent of the tiers.
Item
GPT-6 Astra
GPT-6.1 Sol
GPT-6 Sol
GPT-5.6 Sol
Opus 5.5
Sonnet 5.5
Client
Codex CLI 0.156.1
Codex CLI a
Codex CLI 0.156.1
Codex CLI 0.156.1
Claude Code a
Claude Code a
Reasoning effort
xhigh
xhigh a
xhigh
xhigh
high a
xhigh
Reasoning budget per round
4200 s
4200 s a
4200 s
4200 s
4200 s a
4200 s
Revision window
1800 s
1800 s a
1800 s
1800 s
600 s a
600 s
Selected K1–K17 entries
91
143
255
306
175
175
Counted extra attempts
13
9
35
12
21
40
Table 3: System configurations and evaluation coverage. Each system comprises a model and its runtime configuration, so cross-system comparisons are descriptive. a Settings from session logs; other settings come from configuration files.
System
q0
q17
qmax
τ90
C1
C17
rec.
GPT-6 Astra
0.368
0.842
0.869
9
0.333
0.833
9
GPT-6.1 Sol
0.284
0.689
0.728
7
0.222
0.667
8
GPT-6 Sol
0.031
0.472
0.472
17
0.000
0.389
7
GPT-5.6 Sol
0.012
0.104
0.104
17
0.000
0.056
1
Opus 5.5
0.177
0.744
0.744
15
0.111
0.722
11
Sonnet 5.5
0.124
0.700
0.708
13
0.111
0.611
9
Table 4: Curve summaries for A1–A5 (18 tasks). q0 and q17 are the task-mean scores at K0 and K17, qmax the peak, τ90 the first round reaching 90% of the peak gain, C1 and C17 the fractions of tasks solved by K1 and K17, and “rec.” the number of tasks failed at K1 and solved later.
Task
Astra
6.1 Sol
6 Sol
5.6 Sol
Opus
Sonnet
Arrange Largest Number
1
2
17
–
3
2
Classify Objects by Language
1
1
–
–
4
2
Cover Blocks
1
3
4
17
10
9
Deposit Coin
4
–
–
–
15
–
Fill Pen Holder
3
8
–
–
–
–
Insert Key
–
–
–
–
–
–
Table 5: First successful Harness round on each task. A dash denotes no success within K17.
Group
Benchmark item
Source
Spatial
Object detection (mAP@50)
Roboflow
OCR text localization
Roboflow
Visual grounding
ParseBench
Harness
ctr-optimization
Terminal-Bench 4.0
shadow-relay
Terminal-Bench 4.0
Instruction following
LiveBench
Table 6: Public benchmarks behind the capability profiles. Scores were collected on 15 October 2026; each group averages normalized benchmark ranks with equal weight.
Name
Condition
Workspace
Grader
Images
Actions
Full
Full loop
✓
✓
✓
✓
No-img
No earlier-round images
✓
✓
✗
✓
No-grader
No grader feedback
✓
✗
✓
✓
WS+FB
Workspace and grader only
✓
✓
✗
✗
No-hist
No history
✗
✗
✗
✗
Fixed
Full loop, fixed layout
✓
✓
✓
✓
Table 7: Information from earlier rounds in each ablation condition. Workspace: committed notes, skills and code. Grader: result, score and failure reason after each round, and earlier-round outcome files; step-level success flags in action replies and logs are not part of this channel. Images and actions: observation images and action logs from earlier rounds. Current-round observations, including camera images, are available in every condition.
K
Full
Fixed
No-img
No-grader
WS+FB
No-hist
1
1
1
1
0
0
1
2
2
3
1
2
1
2
3
6
6
3
2
2
1
4
6
8
4
2
7
0
5
5
7
3
2
3
0
6
4
7
5
1
7
3
Table 8: Per-round online success in the Astra ablation. Successes out of 12 chains per condition and round (K1–K6 run in full).
Condition
Semantics
Sources
Succ./ n
Rate
E3-full
Full-loop workspaces; test agent may read prior-round raw records
12
23/60
38.3%
E3-notes
Full-loop workspaces; workspace files only
12
25/60
41.7%
HO-Fixed
Fixed workspaces on new layouts
12
27/60
45.0%
HO-NoImg
No-img workspaces; incomplete
9 (6)
13/35
37.1%
HO-NoGrader
No-grader workspaces; incomplete
6 (1)
5/21
23.8%
Table 9: Frozen held-out test of Astra K6 workspaces. 12 sources × 5 held-out layouts for the complete conditions; HO-NoImg and HO-NoGrader are incomplete and not comparable with the complete conditions.
Recursive self-improvement (RSI) describes a system that improves not only its behavior but also the process by which it improves. Recent digital agents test this idea by rewriting their own code and checking each version with executable tests. Bringing these mechanisms into the physical world requires learning from partial observations through costly trials, while accounting for changing operating conditions and potentially irreversible consequences. We position RSI within robot learning and trace pathways from improving robot capabilities to improving the processes that generate them. We examine how physical experience can strengthen these improvement mechanisms and when the resulting gains can support further self-improvement of the mechanisms themselves. Gains carry forward only when feedback addresses remaining learning needs, the receiving component can use the change, its value survives changes in tasks or hardware, and continued practice remains safe and affordable. Existing studies, concentrated in manipulation, support behavioral improvement, benefits to later learning, and procedure improvement. Some systems also change the improver's knowledge or tools, but the physical studies reviewed here have yet to isolate gains attributable to recursion. We synthesize these findings into evaluation principles and research directions for machines that, trial by trial, safely improve the way they improve.
Meng Luo, Zechen Bai, Shanqing Xu +9
National University of Singapore · Huazhong University of Science and Technology · Shanghai Jiao Tong University +2
Robots are increasingly autonomous in task execution, yet their improvement remains largely human-driven. This study investigates Physical Recursive Self-Improvement (Physical RSI), in which physical experience guides lasting updates to both a robot's capabilities and its improvement procedure, making subsequent improvement more effective across tasks. Physical RSI faces challenges from noisy observations, continuous dynamics, costly trials, and irreversible consequences. We present an L0-L5 framework that differentiates capability-level from meta-level updates and within-task from across-task gains, reserving L5 for independently verified recursive gain. We study existing methods around harness adaptation and model training, alongside safety and control mechanisms that govern execution and successive updates. Our analysis reveals that, despite substantial progress in self-improvement, the surveyed literature does not yet demonstrate Physical RSI as defined by our framework. We identify research priorities in reliable feedback, causal attribution, safety preservation, and benchmarking, outlining a path toward robots that autonomously improve how they learn while ensuring safety. The GitHub repository for Physical RSI is available at https://github.com/SAIL-Research-Lab/Physical-RSI.
Hai Zhou, Fan Yang, Zeyi Chen +27
Shanghai Jiao Tong University · Google · Microsoft +13
Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less.
Peidong Wang, Ying Chang, Zhiming Ma +8
Northeastern University · AgiBot · University of Chinese Academy of Sciences