EndoRSI: Physical-Evidence-Driven Recursive Improvement of Endoscopic Robot Policies

Chi Kit Ng; Yidong Zhang; Tianchun Wu; Siu Hing Lui; Hanlin Zhao; Huxin Gao; Hongliang Ren

cs.ROOct 9, 2026

EndoRSI: Physical-Evidence-Driven Recursive Improvement of Endoscopic Robot Policies

Authors: Chi Kit Ng, Yidong Zhang, Tianchun Wu, Siu Hing Lui, Hanlin Zhao, Huxin Gao, Hongliang Ren

Organizations: Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China

Abstract

Artificial intelligence is beginning to automate scientific discovery, yet physical robot-policy research still relies on scientists and engineers to interpret experimental failures and decide how controllers should be revised. We present EndoRSI (Endoscopic Recursive Self-Improvement), a human-governed, self-improving robotic research system in which the embedded EndoForge coding agent iteratively develops and refines endoscopic robot policies using evidence from physical experiments. Physical execution remains isolated within a safety-constrained runtime under human authorisation. In motorised endoscopy, validation evidence first caused the agent to reject an initial action representation, while subsequent physical experiments falsified an offline-supported insertion strategy by revealing hazardous closed-loop behaviour absent from the source data. Failed physical rollouts were then incorporated into policy revision; insertion success subsequently increased from 2/21 in the first generation to 17/21 in the fourth. The frozen final policy retained competence across visual and hardware shifts and was the best performing autonomous policy across six independent ex-vivo porcine specimens. These results show that physical robot interaction can serve as experimental evidence for agent-driven robot policy research.

Figures & tables

Figure 1: EndoForge connects task definition, policy research, physical execution and evidence-driven revision. Existing tracking and navigation data provide complementary source skills. The coding agent analyses data, results and code and prepares candidate policies. Human-authorised execution occurs through a separate safety-constrained runtime. Physical outcomes and human feedback inform subsequent research and transfer evaluation.

Figure 2: Physical evidence changes subsequent research decisions. a , The research loop connects analysis, hypotheses, candidate implementation, authorised experiments, outcome feedback and revision. b , The coding agent acts offline; a separate runtime executes the fixed policy under human supervision. c , Division of responsibilities among the coding agent, human investigators and runtime. d , The experimental progression from source-skill validation to physical falsification, failure-derived supervision and transfer. Success, failure and emergency-stop outcomes are denoted S, F and E, respectively.

Figure 3: Validation evidence redirects policy design. a , Source tracking and navigation tasks, coding-agent activities and candidate dispositions. b , Held-out image-conditioned reference comparisons. Diamond markers indicate the reported MAE values at the bar endpoints; thin horizontal lines lead to the numerical labels and do not indicate uncertainty intervals. c , Validation MAE during the progression from v1 to direct continuous prediction in version 2 (v2) and temporal ensembling (TE). Validation determined candidate selection; held-out evaluation quantified generalisation. The validation and held-out comparisons use different episodes and therefore different baseline values. Exact split-specific values and metrics are reported in Supplementary Note 3 and Supplementary Tables 7 and 8. MAE is expressed in motor units. ZS-B denotes the dual-policy router and ZS-C the instructed VLA policy.

Figure 4: Physical experimentation falsifies an offline-supported insertion strategy. a , Three pre-defined directions and their dispositions: ZS-A servo gate, dual-policy router (ZS-B) and instructed VLA policy (ZS-C). b , First-lunge times for all 16 ZS-A attempts, a representative forward-command and bending trace, and representative frames illustrating premature advancement without a reactive stop. All 16 attempts failed and required emergency stopping. The median first-lunge time was 0.15 s; the gate required four consecutive low-bending predictions.

Figure 5: Physical evidence supports continued policy improvement. a , Development of the instructed VLA direction into a reactive policy. b , Source-derived supervision used to initialise Gen1; the mixture contains 70% skill retention, 20% composed insertion and 10% stop proxies. c , Physical outcomes and research revisions across Gen1–Gen4, including changes to instructions and execution settings. Gen1-to-Gen2 also included success-imitation action-head fine-tuning. d , Successful actions supply cloning targets; unsuccessful states supply no-forward targets with steering retained. e , Success rates across 21 attempts per generation; the line connects the observed proportions, and each dot denotes one trial outcome (S, F or E). Wilson 95% CIs and exact outcomes are reported in Table 1 ; exact comparisons are in Supplementary Table 12. Generation comparisons include checkpoint, instruction and execution-profile changes and do not isolate their individual effects.

Policy arm	Success / $n$	Success, % (95% CI)	Emergency stop
ZS-A rule composition	0/16	0.0 (0.0–19.4)	16
CNN behaviour cloning	0/21	0.0 (0.0–15.5)	0
Action-chunking transformer	1/21	4.8 (0.8–22.7)	1
Diffusion Policy	2/21	9.5 (2.7–28.9)	0
IBVS	3/21	14.3 (5.0–34.6)	0
Success-only, final iteration	3/21	14.3 (5.0–34.6)	0

Table 1: Physical insertion outcomes in the benchtop development environment. All arms contain 21 attempts except ZS-A, which stopped after 16 consecutive emergency stops. Success proportions use Wilson 95% CIs; ordinary failures and emergency stops remain unsuccessful in every denominator. Exact two-sided Fisher comparisons are reported in Supplementary Table 12.

Figure 6: The evolved policy transfers across deployment shifts and biological tissue. a , Gen4 outcomes across deployment conditions (21 attempts per condition). b , Autonomous outcomes in two ex-vivo cohorts (three specimens per cohort; seven attempts per condition and specimen). Dots show individual outcomes; intervals show Wilson 95% confidence intervals. c , Representative starting, entrance-alignment and lumen or terminal views from successful Gen4, successful baseline and failed trials. Failure sequences do not establish lumen centring. These examples add no trials or biological replicates.

Supplementary Figure 1: Commercial endoscopes and motorised robot configurations. The upper row shows the medical gastroscope, its motorised knob adapter and the assembled ER1 platform (Olympus GIF-Q260J; 9.9-mm insertion diameter). The lower row shows industrial endoscopes and adapter designs used for ER2 (Ailishi S50; 6.0 mm) and ER3 (Ailishi S50; 3.9 mm). Labels (1) and (2) identify the illustrated industrial steering adapters; (3) identifies the shared forward actuator. ER1 supplied tracking data, ER2 supplied navigation data and supported policy development and ex-vivo experiments, and ER3 was used for different-instrument transfer. The original hardware illustration is reproduced without modification. Dimensions are those labelled in the supplied illustration; the arrangement is not a calibrated scale comparison.

Supplementary Figure 2: Ex-vivo evaluation and a successful Gen4 trial. a , ER2 evaluation workflow. Here, zero-shot denotes transfer of the fixed policy to ex-vivo tissue without adaptation. b , Six independent porcine tracheae: specimens 1–3 from the initial block cohort and specimens 4–6 from the interleaved replication cohort; no calibrated scale is implied. c , One successful Gen4 trial on specimen 4 from the interleaved replication cohort. The first two rows sample the first 6.00 s; the next two sample the full recording through 22.80 s. These are two temporal samplings of the same trial, with repeated frames and a reset time axis. Labels give zero-based frame indices and elapsed video times at 20 frames s -1 , not scored traversal times. This qualitative example adds no trials or biological replicates.

Supplementary Figure 3: Representative unsuccessful ex-vivo trials. a , Gen4 on specimen 6 from the interleaved replication cohort. b , Gen1 on specimen 2 from the initial block cohort. In each panel, the first two rows sample the first 6.00 s, and the next two sample the full recording, through 95.15 s in a and 8.05 s in b . Each panel shows one trial with two temporal samplings, repeated frames and a reset time axis. Labels give zero-based frame indices and elapsed video times at 20 frames s -1 . Both trials have unsuccessful canonical outcomes; the terminal views do not establish successful lumen alignment. The images alone do not establish failure mechanism or emergency-stop activation. These qualitative examples add no trials or biological replicates.

Stage	Scientific role	Evidence and decision	Frozen output
1	Audit source skills	Verify image–action correspondence, episode splits, ranges and task statistics	Locked splits, normalisation and audit report
2	Test policy hypotheses	Reject v1 on validation; replace discrete residual actions with continuous chunks and temporal ensembling (TE)	Selected source-policy checkpoint and validation report
3	Operationalise a task hypothesis	Human specifies align-then-advance; agent quantifies source-task support, rejects unsupported navigation-centroid steering and implements three candidates	Cross-task analysis; zero-shot composition A (ZS-A), dual-policy router and vision–language–action (VLA) manifests
4	Falsify physically	Sixteen consecutive ZS-A emergency stops reject discontinuous rule composition	Sealed outcome set and transition to reactive policy
5	Develop reactive insertion	Agent develops its pre-frozen VLA direction after physical falsification and a human requirement for end-to-end stop-capable behaviour	Gen1 checkpoint and manifest
6	Evolve from evidence	Gen1 $\rightarrow$ Gen2 clones success only; failure relabelling first enters the Gen2 $\rightarrow$ Gen3 update	Gen2–Gen4 outcome records, recipes and research ledger

Supplementary Table 1: Implemented evidence-to-policy lifecycle.

Component	Input and operation	Output and authority
Human investigators	Define the task and constraints; reset and authorise each trial	Retain deployment and emergency-stop authority; assign terminal S/F/E outcomes
Coding agent	Audit data and results; implement and revise candidates between trials	Prepare a future checkpoint, manifest, instruction and execution limits for human authorisation
Frozen candidate	One authorised checkpoint, manifest, instruction and limit configuration	Remains fixed throughout the armed trial
Visual processing	Full endoscopic frames pass through preprocessing and asynchronous inference	Semantic action predictions enter an action queue
Execution runtime	The 20-Hz loop selects valid queued actions	Deterministic magnitude, watchdog and slew limits bound the command
Platform adapter	Convert semantic commands to the selected platform’s motor mapping	Send platform-specific packets to serial firmware

Supplementary Table 2: Control flow and safety authority. The coding agent acts between trials and has no camera or motor interface while the robot is armed. S, F and E denote success, ordinary failure and emergency stop.

Platform	Instrument	Dataset or experiment role	Lower controller
ER1	Olympus GIF-Q260J medical gastroscope, 9.9 mm	Tracking source dataset in discarded ex-vivo pig stomach	Four-motor firmware
ER2	Ailishi S50 industrial endoscope, 6.0 mm	Navigation source dataset; development and evolution aperture; low illumination; phantom-configuration shift; both ex-vivo cohorts	Six-channel firmware (four motors are used); M1–M4 active, M5/M6 zero
ER3	Ailishi S50 industrial endoscope, 3.9 mm	Different-instrument transfer arm	Same six-channel firmware (four motors are used) and active-channel pattern as ER2

Supplementary Table 3: Instrument, experiment and controller mapping.

Layer	Input and output	Reproducible role in the archive
Platform collection	Camera, timestamp and executed device command $\rightarrow$ recorded episode	Three device-facing collection paths preserve the raw embodiment and acquisition context
Dataset conversion	Recorded episode $\rightarrow$ aligned image–semantic-action records	Validates image/action correspondence, constructs episode-disjoint splits and converts commands to $[m_{1},m_{3},m_{4}]$
Training	Canonical records and task normalisation $\rightarrow$ frozen checkpoint and manifest	Learns the action controller without embedding a raw ER1/ER2/ER3 serial packet
Inference runtime	Current image, instruction and frozen manifest $\rightarrow$ bounded semantic action	Runs asynchronous model inference and a deterministic 20-Hz watchdog/safety boundary
Hardware adapter	Semantic action plus platform profile $\rightarrow$ device motor packet	Applies channel expansion, sign, scale, gain, bias and inactive-channel rules
Lower firmware	Device motor packet $\rightarrow$ motor actuation	Separate four-motor medical and six-channel industrial Keil projects

Supplementary Table 4: Relationship among collection data, learned policy and hardware code.

Task	Episodes	Action rows	Split	Test $n$	Collection system and environment
Tracking	100	12,710	80/10/10	10 episodes	ER1, Olympus GIF-Q260J; ex-vivo pig stomach
Navigation	100	16,534	80/10/10	10 episodes	ER2, Ailishi S50 6.0 mm; colon phantom

Supplementary Table 5: Source-dataset audit and experimental units.

	Tracking		Navigation
Policy	Validation	Held-out	Validation	Held-out
CNN behaviour cloning	2.619	2.316	19.463	23.580
Agent v1	2.654	2.065	19.811	23.608
Agent v2	2.615	2.255	19.091	22.032
Agent v2+TE	2.426	2.112	18.791	21.683

Supplementary Table 6: Validation-based model selection and post-selection held-out performance. Values are frame-pooled mean absolute error (MAE) in raw motor units; lower is better. Each task contains 100 episodes split into 80 training, 10 validation and 10 test episodes. Validation determined candidate selection; held-out scores were evaluated afterwards. TE denotes causal temporal ensembling of the selected v2 checkpoint and is a decoding procedure, not a separately trained model.

Task	Policy	MAE	RMSE	Sign accuracy
Tracking	Classical visual servoing	2.846	4.006	0.778
	Diffusion Policy	2.340	3.891	0.780
	Action-chunking transformer	2.417	3.830	0.781
	CNN behaviour cloning	2.316	3.697	0.817
	Image-only VLA	2.622	4.223	0.611
	v1	2.065	3.884	0.741

Supplementary Table 7: Complete held-out image-conditioned offline results. Lower MAE and RMSE are better; higher sign accuracy is better. The copy-previous diagnostic is excluded because it is teacher forced.

Task	Original	Seed 1	Seed 2	Mean $\pm$ s.d.	CNN reference
Tracking	2.112	2.076	1.995	$2.061\pm 0.060$	2.316
Navigation	21.683	21.257	21.615	$21.518\pm 0.229$	23.580

Supplementary Table 8: Selected-policy stability across three training runs. Values are held-out test MAE; s.d. denotes standard deviation.

Claim or candidate	Sample size	Evidence	Decision
Tracking centre error predicts bending	10,284 frames	$R^{2}=0.340$ ( $m_{3}$ ), 0.370 ( $m_{4}$ )	Retain as alignment primitive
Navigation dark-centroid error predicts bending	13,810 frames	$R^{2}=-0.034$ ( $m_{3}$ ), 0.004 ( $m_{4}$ )	Reject
Low bending predicts advancement	13,810 frames	Mean $m_{1}=16.746$ in lowest quartile, $\leq 0.560$ otherwise	Retain as conditional advance principle
ZS-A composes both rules safely	16 physical attempts	0 success; 16 emergency stops	Physical falsification rejects rule composition

Supplementary Table 9: Source-task evidence and its disposition.

Update	Source-batch outcomes	Training use	Successor outcome
Gen1 $\rightarrow$ Gen2	2 success; 19 failure/emergency stop	Success only; 175 cloning samples; 300 steps; loss 0.4662 $\rightarrow$ 0.2851	8/21
Gen2 $\rightarrow$ Gen3	8 success; 13 failure/emergency stop	Failure-derived: 68 cloning + 144 mirrored avoidance samples from 12 contributing records; nine records supplied no sample; loss 0.2243 $\rightarrow$ 0.0918	14/21
Gen3 $\rightarrow$ Gen4	14 success; 7 failure	Failure-derived: 240 cloning + 160 avoidance samples; 300 steps; loss 0.0914 $\rightarrow$ 0.0643	17/21
Success-only control	1/21 $\rightarrow$ 3/21 $\rightarrow$ 3/21 $\rightarrow$ 3/21	Each update used successes from its own preceding batch and discarded unsuccessful outcomes	3/21 final

Supplementary Table 10: Generation-specific development evidence and update recipes. Source-batch outcomes are recorded in the supplied per-trial outcome correspondence; the last column gives the final video-reviewed performance of the resulting policy. Comparisons between lineages are descriptive because their evidence pools and starting checkpoints differed.

Group	Arm	Success / $n$	Success, % (95% CI)	Emergency stop
Physical falsification	ZS-A	0/16	0.0 (0.0–19.4)	16
Fixed reference	CNN behaviour cloning	0/21	0.0 (0.0–15.5)	0
	Action-chunking transformer	1/21	4.8 (0.8–22.7)	1
	Diffusion Policy	2/21	9.5 (2.7–28.9)	0
	IBVS	3/21	14.3 (5.0–34.6)	0
Success-only	Iteration 1	1/21	4.8 (0.8–22.7)	0

Supplementary Table 11: Complete benchtop outcomes with exact sample sizes and Wilson 95% confidence intervals (CIs). Emergency-stop counts come from benchtop operator 1’s live record; endoscopic video alone does not establish emergency-stop activation.

Comparison	Odds ratio	$P$	Role
Gen4 versus Gen1	40.38	$5.75\times 10^{-6}$	Overall generation contrast
Gen4 versus IBVS	25.50	$3.21\times 10^{-5}$	Best fixed classical reference
Gen4 versus final Success-only iteration	25.50	$3.21\times 10^{-5}$	Failure-evidence mechanism control
Gen2 versus Gen1	5.85	0.0670	Generation comparison
Gen3 versus Gen2	3.25	0.121	Generation comparison
Gen4 versus Gen3	2.13	0.484	Generation comparison

Supplementary Table 12: Two-sided Fisher exact tests for final phantom contrasts and descriptive generation comparisons. Generation comparisons quantify observed proportions; causal interpretation is limited because the Success-only and generation lineages used different evidence pools and starting checkpoints.

Human investigators	Coding agent
Define research goal, align-then-advance task and bounded deployment authority.	Plan dataset audit, validation-based revision and physical evaluation.
Provide existing tracking and navigation datasets and experimental constraints.	Audit image-action records, lock episode splits and implement comparison policies.
Require validation evidence to guide policy improvement.	Reject v1 on validation. Implement continuous action chunks and causal temporal ensembling.
Specify insertion as tracking plus navigation. Require freezing before insertion outcomes.	Quantify cross-task support. Reject centroid steering. Freeze ZS-A, ZS-B and instructed VLA.
Authorise ZS-A tests and use emergency stop in all 16 attempts.	Analyse discontinuous advancement and absence of a reactive stop response.
Prioritise end-to-end VLA development and specify stop-on-contact behaviour.	Develop the previously frozen VLA direction into Gen1 using source-derived supervision.

Supplementary Table 13: Human–agent division of scientific decisions. Rows follow the recorded order of research decisions. Investigators retain deployment and emergency-stop authority. ZS-B denotes the dual-policy router.

Policy	Specimen	$n$	S	F	E	Success navigation time, s
IBVS	1	7	3	3	1	6.2 [4.9–14.6]
	2	7	3	3	1	2.0 [1.9–2.1]
	3	7	3	4	0	2.5 [1.9–3.2]
Gen1	1	7	2	5	0	9.3 [8.9–9.6]
	2	7	4	3	0	3.5 [2.4–4.5]
	3	7	3	3	1	6.7 [4.7–6.7]

Supplementary Table 14: Initial-cohort ex-vivo outcomes and successful navigation times by policy and specimen. The specimen is the biological unit; attempts are repeated technical measurements. F and E are distinct live-operator events but both are binary failures. Navigation time is conditional on success and shown as median [IQR].

Condition	Success / $n$	Success, % (95% CI)	Exploratory $P$ vs Gen4	Scope
Initial block cohort
Gen4	21/21	100.0 (84.5–100.0)	–	Frozen evolved policy
Success-only	11/21	52.4 (32.4–71.7)	$4.79\times 10^{-4}$	Final success-only control
Gen1	9/21	42.9 (24.5–63.5)	$5.32\times 10^{-5}$	Initial reactive policy
IBVS	9/21	42.9 (24.5–63.5)	$5.32\times 10^{-5}$	Classical reference
Interleaved replication cohort

Supplementary Table 15: Cohort-specific pooled descriptive outcomes and exploratory attempt-level comparisons. Each cohort contains three independent specimens and seven technical repetitions per condition–specimen cell. Cohorts are analysed separately. Wilson 95% CIs are descriptive; two-sided Fisher tests compare autonomous controls with Gen4 within the same cohort and do not provide specimen-level or causal inference. The initial Gen4 row includes its 21 protocol attempts only. Teleoperation rows are single-operator references and were not tested for an experience effect.

Arm	Outcome	$n$	Mean, s	Median [IQR], s	Range, s	Emergency stop
Success-only 1	F	20	4.6	4.1 [3.6–5.0]	3.3–10.2	0
Success-only 2	F	18	3.5	3.2 [3.1–3.7]	2.5–7.0	0
Success-only 3	F	18	3.8	3.5 [3.2–4.2]	3.0–5.6	0
Success-only 4	F	18	4.8	4.2 [3.5–4.6]	2.7–13.9	0
Gen1	F+E	19	12.4	5.3 [3.8–6.9]	1.9–97.8	5
Gen2	F	13	9.2	8.0 [6.9–9.9]	3.4–21.2	0

Supplementary Table 16: Descriptive video-duration phenotypes for Success-only and generation arms. Duration includes the recorded trial video and is influenced by operator termination; it is not a time-to-event efficacy analysis. Median values are shown with IQR. Emergency-stop counts come only from benchtop operator 1’s live record.

Domain	Comparison	Exact-path $n$	Agreement, %	Reliability
All scored trials	Relevant live record vs external evaluator 1	546	100.0	Cohen’s $\kappa=1.00$
All scored trials	Three rating sources per trial	546	100.0	Fleiss’ $\kappa=1.00$
Phantom	Benchtop operator 1 vs video evaluator 1	315	100.0	Cohen’s $\kappa=1.00$
Phantom	Benchtop operator 1 vs external evaluator 1	315	100.0	Cohen’s $\kappa=1.00$
Phantom	All three rating sources	315	100.0	Fleiss’ $\kappa=1.00$
Initial ex vivo	Initial ex-vivo operator 1 vs video evaluator 2	84	100.0	Cohen’s $\kappa=1.00$

Supplementary Table 17: Binary outcome agreement after analysing every E outcome as F. The scored set comprises 315 phantom trials, 84 initial-cohort ex-vivo trials and 147 replication-cohort trials. The three-way analysis uses three rating sources per trial: live record, video evaluator 1 and external evaluator 1 for benchtop trials; live record, video evaluator 2 and external evaluator 1 for the initial cohort; and the three video evaluators for replication. Emergency-stop counts are not inferred from video and therefore come only from the relevant live record. Exact-path matching is used, with complete binary outcome ratings from the three selected sources for all scored trials.

Cohort	Analysis	$n$	Estimate	95% CI	Interpretation
Initial block	ICC(A,1)	48	0.994	0.978–0.998	Individual-rating agreement
Initial block	ICC(A, $k$ )	48	0.998	0.993–0.999	Three-rating mean reliability
Interleaved replication	ICC(A,1)	76	0.956	0.873–0.984	Individual-rating agreement
Interleaved replication	ICC(A, $k$ )	76	0.985	0.954–0.995	Three-rating mean reliability

Supplementary Table 18: Reliability of successful ex-vivo navigation-time annotations. The intraclass correlation coefficient (ICC) is reported as ICC(A,1) for two-way random-effects absolute agreement of a single rating and as ICC(A,

k

) for reliability of the three-rating mean. ICC 95% CIs use 10,000 trial-level bootstrap resamples with a fixed random seed for each cohort.

Initial block cohort
Protocol condition	Specimen 1	Specimen 2	Specimen 3	Total
IBVS	3/7	3/7	3/7	9/21
Gen1	2/7	4/7	3/7	9/21
Success-only, original configuration	3/7	4/7	4/7	11/21
Gen4	7/7	7/7	7/7	21/21
Interleaved replication cohort

Supplementary Table 19: Transfer outcomes across the two ex-vivo cohorts and six porcine tracheae. Specimens are independent biological units; each specimen cell contains seven technical repetitions, and Total aggregates the three cells within a cohort. Each cohort block lists only conditions evaluated in that protocol. Within replication, the two Success-only rows use the same checkpoint and differ only in the runtime execution profile: action scaling, gains and smoothing (Supplementary Note 5).

Explore similar work

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation
Sep 30, 2026Haoran Lang, Haotao Lu, Shiyu Sang +6Robot Policy Adaptation Robotic Data Collection
InstAdapt · ShanghaiTech University
ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
Jun 18, 2026Wenli Xiao, Jia Xie, Tonghe Zhang +14Robot Policy Learning Robot Skill Learning
1NVIDIA · 2CMU · 3UC Berkeley
PhysEvo: Astra Can Act, Let It
Oct 6, 2026Wenqing Tian, Zeyu Zhang, Zhaocheng Liu +3Self-Evolving Agents Robot Policy Adaptation
University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · Independent Researcher +1

Sep 30, 2026J/K move · Enter open · S save

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation

Haoran Lang, Haotao Lu, Shiyu Sang, Haoyang Luo+5

InstAdapt · ShanghaiTech University

Robot Policy Adaptation Robotic Data Collection Robotics Simulation Sim-to-Real Transfer Data-Efficient Robot Learning Recursive Self-Improvement

Open paper AskPDF arXiv

Adapting robot manipulation policies to new tasks and environments remains highly data-intensive, while the data needed for further improvement depends on the policy's current capabilities and failure modes. We introduce EmbodiRSI, an agentic system for recursive self-improvement (RSI) in a real-to-sim-to-real setting, where task-specific simulations are constructed from target deployment scenarios and used as low-cost environments for iterative policy improvement before transfer back to the physical world. EmbodiRSI uses policy execution feedback to guide subsequent experience acquisition and policy updates. Two complementary mechanisms close this loop: Collaborative Error Correction generates agent-assisted corrective trajectories from policy-reached states, while Adaptive Data Collection directs expert demonstration generation toward the current policy's weaknesses. The task-specific simulation serves as a reusable workspace for policy warm-up, repeatable evaluation, failure diagnosis, and targeted data generation across successive RSI rounds. Across three tabletop environments and 14 subtasks, EmbodiRSI increases scene-balanced autonomous simulation success from 50.4% to 83.5% over two RSI updates. With 400 adaptive simulated trajectories and only ten real-world refinement trajectories per subtask, EmbodiRSI achieves 83.1% scene-balanced autonomous real-world success, compared with 75.0% for adaptation using 200 real-world demonstrations per subtask. These results demonstrate that feedback-driven recursive improvement in deployment-specific simulations can enable data-efficient adaptation of embodied policies to physical environments.

Abstract

Explore similar work

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation

ENPIRE: Agentic Robot Policy Self-Improvement in the Real World

PhysEvo: Astra Can Act, Let It