Organizations: School of Artificial Intelligence, University of Chinese Academy of Sciences, China · The State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, China · Southern University of Science and Technology, China · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China · Key Laboratory of Computing Power Network and Information Security, Ministry of Education; Shandong Computer Science Center, Qilu University of Technology (Shandong Academy of Sciences), China · Shandong Provincial Key Laboratory of Computing Power Internet and Service Computing, Shandong Fundamental Research Center for Computer Science, China
Embodied systems can update policies, world models, skills, and executable programs from deployment experience, allowing one interaction to benefit later tasks. When a system also revises and reuses its data-acquisition, feedback, or learning procedures, experience changes the process that produces capability, introducing recursive self-improvement. This paper surveys research in physical AI through the connection between persistent capability and its improvement process. We first explain the functional distinction between learner updates and improvers, then compare policy learning, imagined experience, memory and skill accumulation, and executable harnesses. These approaches connect to revisions of curricula, rewards, and research programs in different ways. Contact validity, recovery, retention, and resource costs determine whether the resulting gains accumulate during continued deployment. An L0-L3 rubric distinguishes task-local adaptation, persistent capability gain, inherited process revision, and verified improver gain. Representative benchmarks and within-study contrasts relate these distinctions to existing results. We further propose common-data and interactive comparisons that evaluate reference and revised improvers from shared learner starts, separating the use of experience from its acquisition. This perspective connects the diverse methods, physical operating conditions, and evaluation evidence, underpinning embodied RSI. A comprehensive list of studies in this survey is available at https://github.com/nkd-lkz/physical-ai-rsi-survey.
Figures & tables
Fig. 1: Overview of the review framework for recursive self-improvement in physical AI. (a) Conceptual distinctions among programmed robots, learned policies, persistent self-improvement, and recursive improvement. (b) Physical experience and feedback support capability updates, while system-generated process revisions are retained and reused to govern later learning. (c) Persistent artifacts can modify the learner, the improver, or both. (d) The evidence rubric distinguishes task-local adaptation (L0), persistent capability gain (L1), inherited process revision (L2), and verified improver gain (L3). L3 requires greater subsequent gains from shared learner starts under matched resource budgets and independent tests. The levels specify evidence requirements for individual claims. Comparisons are schematic, and robot scenes are illustrative renderings.
Survey
Organizing focus
Connection to this paper
Self-evolving Embodied AI [ 10 ]
Memory, tasks, world models, embodiment, and model evolution
Broad embodied framework, including changes to optimization and evaluation procedures.
Weights or Skills? [ 11 ]
Execution feedback, persistent memory, and code-based skill search
Operational distinctions between task-local repair and reusable capability.
Self-Evolving AI for Humanoids [ 13 ]
Post-deployment mechanisms, meta-updates, safety, and lifetime trajectories
Physical evolution and trajectory evaluation, complementing separate tests of process inheritance and improver gain.
The Last AI Built by Humans [ 14 ]
RSI autonomy levels, application domains, and controlled improver evaluation
Direct RSI perspective; autonomy and evidential strength characterize different aspects of improvement.
Continual Learning for Robotics [ 12 ]
Continual-learning strategies, retention, transfer, and robotics constraints
Foundations for persistent-capability assessment and physical evaluation conditions.
TABLE I: Relation to closely related surveys. Comparisons concern organizing perspectives and overlapping coverage.
Fig. 2: Information boundaries between physical interaction, learner updates, and improver revision. Observations supply experience for learner updates; system-generated procedure changes are retained and used in later update rounds. Development screening can inform these revisions. A concealed reporting audit evaluates frozen candidates on independent held-out outcomes and supplies no feedback to the updates being assessed. Grids, contact geometry, records, and program differences are schematic representations.
Level
Operational claim
Minimum distinguishing evidence
L0
Task-local adaptation
Feedback changes the current attempt or episode; the adapted state is discarded at the declared reset.
L1
Persistent capability gain
Experience produces retained capability and benefits later independent attempts or tasks.
L2
Inherited process revision
System-generated revisions of an improvement component persist and govern subsequent learner updates.
L3
Verified improver gain
The revised improver yields greater subsequent gains from shared learner starts under matched resource limits and independent tests.
TABLE II: Progressive claims and minimum evidence
Fig. 3: Evidence requirements across a declared reset boundary. Episode context c, retained learner state x, and improvement procedure φ have different persistence roles. L0 describes task-local changes that reset; L1 concerns retained capability gains in later independent attempts. L2 requires a system-generated process revision to be inherited and used in a later learner update. L3 requires greater subsequent gains under shared learner starts, matched resource budgets, and independent tests. The diagrams illustrate the criteria and corresponding evaluation conditions.
Fig. 4: Developmental trajectories of self-improvement in Physical AI. The 46 selected works are grouped by mechanism, with a separate column for recent work. Branches indicate thematic grouping, not causal inheritance or increasing RSI maturity. Intervals have unequal duration; within-period placement follows mechanism. R, S, and D denote reported real-robot, simulated/virtual, and non-embodied settings, not evidence strength. P denotes a project report without an assigned experimental-setting tag. Persistent updates and task gains alone do not establish improver gain.
Fig. 5: Four carriers of persistent physical experience. Policy and value updates change parameters and action selection; world models retain predictive structure learned from observed transitions; indexed memory and skill libraries support retrieval and composition; programs and harnesses retain changes to execution logic, illustrated by a recovery branch. Each panel exposes the retained representation and its later use. A program change is classified by its functional role in execution or in producing later updates.
Fig. 6: Fixed-model planning and model–policy co-improvement. The fixed-model planner evaluates imagined action alternatives and executes a selected action while retaining model parameters. In the coupled loop, real transitions train the model, the updated model generates imagined trajectories for policy learning, and the updated policy collects further physical experience. Solid and dashed paths distinguish physical and model-generated experience. The observed–predicted offset is schematic; coupling alone does not establish inherited revision of the improvement procedure.
Fig. 7: Four locations for inherited improver revision. Changes to acquisition alter the selected experience; changes to feedback alter the training signal; changes to learning or search alter the update routine; changes to a development gate alter which candidates are retained. Highlighted differences are accepted, inherited, and used in a later cycle. The rows describe alternative intervention locations. Development-gate revisions remain subject to declared external constraints, while independent reporting tests stay fixed for the comparison.
Fig. 8: Physical continuity and learning-state retention are separate, coordinated processes. Contact changes or calibration drift can require physical recovery before redeployment. Candidate changes to retained state are checked on earlier and new tasks, then accepted or rolled back to the previous checkpoint. Recovery is not itself a learning update. The resource ledger spans collection, recovery, updating, and verification, including robot time, computation, human effort, and wear or repair.
Fig. 9: Conceptual physical failure modes. (a) A plausible pose can yield unstable contact, illustrated by contact-normal and tangential-slip directions. (b) Partial observations can produce a false-positive judgment despite an unsuccessful physical outcome. (c) A new-task update can disrupt an earlier skill. (d) Sensor calibration or actuator offsets can change the observation–action mapping. The scenes, state changes, and offsets are illustrative rather than experimental measurements.
Field
What must be recorded
Persistent change
Weights, world model, memory, skill code, harness, reward, or update procedure; record functional role.
Feedback and setting
Physical measurements, human labels, learned judge, simulator, or hybrid; distinguish physical and digital evidence.
Retention boundary
Across attempts, independent episodes, tasks, or deployments; specify every reset.
Human and resource input
Demonstrations, resets, approval, maintenance, corrections, compute, time, and discarded trials.
Claim and control
Capability gain; inherited process revision; improver gain; shared starts, fixed references, and evaluator independence.
TABLE III: Evidence fields used independently of the L0–L3 rubric
Method
Comparison object
Evaluation focus
Zeva [ 23 ]
Task context
Memory boundaries across episode resets.
MemBodied [ 103 ]
Episode memory
Within-episode memory and persistent benefit across episodes.
RECAP [ 2 ]
Policy updates
Retained capability gains under a fixed learning procedure.
Learning While Deploying [ 3 ]
Fleet policy learning
Effects of fleet experience and model versions.
RISE [ 49 ]
Model-assisted policy learning
Training gains and the contribution of planning.
Motus2 [ 53 ]
Policy and planning
Policy retention and planning gains under shared initialization.
TABLE IV: Improvement objects and evaluation focus of representative methods
Fig. 10: Evaluation objects in representative robot-learning protocols. Cells name the task structure, split, or metric targeted by each scoped protocol. Dashes denote objectives not encoded as direct targets in that protocol, rather than capabilities absent from the environment. CALVIN instruction chains differ from a learning history; LIBERO and Continual World explicitly evaluate sequential retention and transfer. The RoboCasa365 row denotes its public multi-task leaderboard. Sources: RLBench, Meta-World, ManiSkill2, CALVIN, LIBERO, Continual World, RoboCasa, and RoboCasa365. [ 197 , 196 , 198 , 205 , 201 , 183 , 202 , 206 ]
Fig. 11: Common-data and interactive conditions for improver comparison. In the common-data condition, cloned learner starts and identical permitted evidence enter the reference and revised improvers. In the interactive condition, each improver acquires its own evidence under matched starts and resource ceilings. Branch-specific successor learners are assessed by independent held-out tests that provide no update feedback. The two conditions assess evidence use and acquisition plus updating, respectively; report subsequent gain, retention, transfer, and actual costs.
Fig. 12: Crossed evaluation of learner starts and improvement procedures. Each row clones a learner start and applies the reference and revised improvers; the horizontal contrast uses G and ΔG as defined in Eq. ( 3 ), with common budgets and test conditions omitted from the displayed notation. Diagonal comparisons change both learner state and improver. Estimate uncertainty over independent replicates and report absolute gains alongside relative contrasts. Match resource ceilings and record realized consumption, including construction, collection, computation, human assistance, energy, and resets. The diagram contains no measured results.
Control
Purpose and required report
Shared learner starts
Match model version, initial weights, retained skills, prior data, and accessible tools.
Common-data condition
Give both improvers identical admissible evidence; isolate how the evidence is used.
Interactive condition
Allow separate collection under matched physical and compute limits; count unsuccessful acquisition.
Independent assessment
Separate development feedback from held-out outcomes; freeze the assessor or calibrate versions externally.
Retention and transfer
Test prior tasks, new tasks, and compatible alternative learners; report negative transfer.
Repeated runs and costs
Report uncertainty, full trajectories, total search cost, human assistance, resets, and maintenance.
TABLE V: Controls for an improver-gain claim
Transfer
Occlusion
Counting
Sequence
Overall
Method
TSR
CSR
TSR
CSR
TSR
CSR
TSR
CSR
TSR
CSR
π0.5
20.0
42.8
12.7
17.2
14.3
50.9
60.0
71.6
21.5
38.7
HiF-VLA
17.5
38.9
12.7
27.1
8.6
45.9
42.5
70.2
16.9
39.8
MemoryVLA
15.0
37.2
7.3
13.1
14.3
55.1
37.5
65.2
15.0
35.3
MemER
20.0
36.1
16.4
33.2
27.1
65.1
65.0
79.1
27.3
49.1
PrediMem [ 223 ]
22.5
45.2
27.3
38.4
45.7
69.3
72.5
89.5
38.5
55.2
TABLE VI: Reported task and cumulative success on the full 26-task RoboMemArena benchmark (%).
Category score
Average
Method
Type
Gen-S
Gen-R
Prec.
Long
Mem.
Open
Score
SR (%)
PhysicalRSI 1.0
Agent+VLA
31.57
11.04
38.05
46.37
46.74
28.88
36.27
31.38
Simate-beta
VLA
40.54
29.63
34.35
57.84
33.33
9.12
33.95
27.96
VPP2-Preview
WAM
43.92
18.75
33.75
35.31
50.78
5.83
31.40
25.62
Liber-0 Preview
VLM+WAM
39.31
10.67
38.28
45.98
37.77
6.68
30.74
25.52
GPT-6 Astra
Agent
35.32
31.40
12.65
21.45
43.04
34.36
28.97
22.48
TABLE VII: Reported RoboDojo simulation results from the 28 September 2026 leaderboard snapshot.
Method
Coverage
Retained artifact
Revision and testing
Evidence focus
PhysicalRSI 1.0 [ 62 ]
Simulation
Harness and skills
Candidate revisions are tested, selected and inherited.
System evolution and task performance.
HexaAnything [ 226 ]
3 tasks
Tool code
Two revision rounds; fixed evaluator. Cloth trials include human-selected action points.
Direct tool-revision trajectory.
URAI [ 227 ]
5 tasks
Tool code
Human-guided development; tools frozen for four execution models.
Tool use and a separate revision case.
ENCORE [ 228 ]
22 tasks
Policy programs
Demonstration-guided refinement; programs frozen before sealed tests.
Program discovery and generalization.
RoboHarn-Evo [ 106 ]
2 tasks
Physical knowledge
Knowledge evolves on RMBench and remains read-only on RoboDojo.
Transfer of evolved knowledge.
ME-Brain-1.0 [ 229 ]
54 configurations
Memory and skills
RoboDojo evaluates the action model; multi-round evolution is not isolated.
Task performance of an evolvable system.
TABLE VIII: Improvement mechanisms and evaluation scope of RoboDojo studies.
Method
Evaluated scope
Changed component
Reference
Updated
Δ (pp)
EmbodiRSI [ 46 ]
S1+S3 simulation
Data acquisition
59.3
79.9
+20.6
FIND [ 84 ]
8 Franka tasks
Residual policy
55.0
71.9
+16.9
RoboCoach [ 54 ]
3 Franka tasks
Skill policies
13.3
75.0
+61.7
Motus2 [ 53 ]
2 real-robot tasks
Action parameters
65.0
72.5
+7.5
LWD [ 3 ]
Freezer restocking
Fleet policy
92.0
97.0
+5.0
RPG [ 230 ]
4 simulation tasks
Skill library
65.0
92.5
+27.5
TABLE IX: Within-study performance changes after improvement.
Method
Reference and update setting
Experience or resource budget
Evaluation
Q-VGM
Few-shot SFT; offline then online
150 offline rollouts/suite
500 tests/suite; 3 training seeds
RedFlow
Base policy; one offline update
1,536 rollouts/suite
500 tests/suite; horizons 240/240/240/480
Q-Planning
Frozen FastWAM; learned Q
100 rollouts/task/round; 10 rounds
20 tests/task; horizon 540 for every row
WoVR
Full-trajectory SFT; imagined RL
2,500 simulator trajectories/suite
Four LIBERO suites
EmbodiRSI
Static versus adaptive acquisition
400 retained trajectories/subtask at each endpoint
S1+S3 scene-balanced SR; no HIL
FIND
Base versus online residual policy
50 initial demonstrations/task
5 runs; 20 independent tests/task/checkpoint
TABLE X: Data, update conditions and evaluation resources for the reported contrasts.
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
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.