PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.13376v1cs.RO

TL;DR

Robot-policy research still commonly depends on people to interpret experimental failures and decide how controllers should be revised. EndoRSI uses a human-governed coding agent to refine endoscopic policies from physical experiments, and the final policy retained competence across visual and hardware shifts and performed best among autonomous conditions on every tested porcine specimen.

  • Problem

    Robot-policy development still rarely automates the interpretation of physical evidence, policy revision, and retesting, while researchers commonly decide how poor results should change controllers.

    Source: Introduction · pp. 1–2

  • Method

    EndoRSI uses a human-governed coding agent to iteratively develop endoscopic robot policies from physical-experiment evidence, while a separate safety-constrained runtime executes only human-authorised policies.

    Source: Introduction · p. 2, Results · p. 3

  • Results

    The final policy retained competence across visual and hardware shifts and was the highest-performing autonomous condition on every discarded porcine specimen.

    Source: Discussion · p. 11

  • Takeaways & Limitations

    Physical robot experiments can serve as evidence that changes subsequent AI-driven policy development, including what the agent believes and builds.

    Source: Discussion · pp. 10–11

    The comparison of policy lineages does not isolate a causal effect of failure supervision because they differed in positive-evidence pools, checkpoints, instructions, and execution settings.

    Source: Physical evidence supports continued… · p. 6

Abstract

from arXiv · show

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.

Introduction · Results

EndoRSI uses physical-experiment evidence to revise endoscopic robot policies within a human-authorised, safety-constrained research loop. The policy improved across generations and retained competence across deployment shifts and ex-vivo specimens, while physical tests exposed limits that offline evidence missed.

Introduction

  • EndoRSI addresses the uncommon automation of physical-evidence interpretation and policy revision by pairing an iterative coding agent with human authority over deployment and emergency intervention.The agent analyses outcomes and revises policies, while physical execution remains isolated in a separate safety-constrained runtime requiring human authorisation.Source: Introduction · pp. 1–2

Results

  • The Results evaluate four stages: validation-driven redesign, physical falsification of an insertion strategy, failed-rollout supervision and transfer of a frozen evolved policy.The coding agent works at the research timescale; only human-authorised policies execute in the separate safety-constrained runtime.Source: Results · p. 3

Validation evidence redirects policy development

  • Held-out v2+TE reduced tracking MAE from 2.316 to 2.112 and navigation MAE from 23.580 to 21.683 versus CNN references across three independent training runs.Validation evidence led the agent to replace v1’s discrete-bin-plus-residual action representation with direct continuous prediction and temporal ensembling.Source: Validation evidence redirects policy… · p. 3, Figure 3 · p. 5
  • A teacher-forced copy-previous diagnostic achieved lower error while ignoring images, showing that offline error could not establish competence in robot-created closed-loop states.The demonstrations had strong temporal autocorrelation, so validation could reject an inferior representation without proving autonomous physical competence.Source: Validation evidence redirects policy… · p. 5

Physical interaction falsifies policy composition

  • 0/16 ZS-A physical attempts succeeded, and all required emergency stopping; the median first lunge occurred at 0.15 s, exposing hazardous closed-loop behaviour absent from offline evidence.A discontinuous composition of alignment and advancement lacked a contact-conditioned stop response; testing ended after 16 consecutive emergency stops. The agent then developed the instructed VLA candidate into Gen1 with conservative stop proxies and a positive forward-slew limit.Source: Physical interaction falsifies policy… · pp. 5–6

Physical evidence supports continued policy improvement

  • 2/21 Gen1 successes rose to 8/21 for Gen2, 14/21 for Gen3 and 17/21 for Gen4, while the Success-only lineage plateaued at 3/21.Failed rollouts contributed conservative no-forward targets with recorded steering retained; the comparison does not isolate a causal effect of failure supervision because the lineages also differed in data pools, checkpoints, instructions and execution settings.Source: Physical evidence supports continued… · p. 6, Figure 5 · p. 8

Policy transfer across deployment shifts and tissue

  • Gen4 retained 13–15 successes in 21 attempts under reduced illumination, a phantom-configuration shift and a different endoscope, with performance declining but not collapsing.The policy was evaluated without adaptation.Source: Policy transfer across deployment… · p. 8, Figure 6 · p. 9
  • 16/21 insertions succeeded with Gen4 in the second ex-vivo cohort, versus 2–7/21 for autonomous controls; Gen4 was the highest-performing autonomous policy on all six specimens.Experienced and novice teleoperation references each achieved 21/21. The six independent specimens were the biological units; repeated attempts were technical repetitions.Source: Policy transfer across deployment… · pp. 8, 10, Figure 6 · p. 9

Discussion

EndoForge shows that physical experiments can change what an AI-driven policy agent believes and builds, rather than serving only as performance evaluations. The system remains human-governed, with physical execution under human control and safety constraints.

Discussion

  • Validation led EndoForge to abandon an inferior action representation, physical deployment rejected a plausible skill composition, and failed rollouts changed later policy supervision.The central contribution is the sequence in which physical evidence changed what the agent believed and built, not the final success rate alone.Source: Discussion · pp. 10–11
  • Unlike traditional imitation learning, reinforcement learning, and online adaptation [34] [38] [39], EndoForge revises representations, policy compositions, and supervision above optimization, where physical closed-loop behavior can invalidate source-data-supported hypotheses.The proposed evaluation asks whether experiments can reject a fixed candidate, change the next implementation, make failures informative, and support transfer beyond development conditions.Source: Discussion · p. 11
  • Physical testing falsified ZS-A’s source-data-supported align-then-advance composition by exposing hazardous closed-loop behavior absent from offline replay.Human investigators supplied the decomposition, and offline replay confirmed that its gate behaved as designed; physical embodiment nevertheless rejected the hypothesis.Source: Discussion · p. 11
  • Continued generational gains followed conversion of hazardous terminal states into avoidance targets, while the Success-only control plateaued.Instruction changes and the Gen1-to-Gen2 execution-profile change prevent attributing the gains specifically to failure supervision or action-head updates. Applying Gen4 runtime settings to the Success-only checkpoint recovered part, but not all, of the gap.Source: Discussion · p. 11
  • Gen4 was the highest-performing autonomous condition on every discarded porcine specimen, with its advantage reproduced in three more; the 16/21 replication still trailed perfect teleoperation references.The relative advantage was reproduced under an interleaved protocol with expanded autonomous and teleoperation controls. The final policy transferred without adaptation to two ex-vivo cohorts; the evidence covers six biological specimens and single-operator teleoperation references.Source: Discussion · p. 11
  • EndoForge is not a fully autonomous scientist: investigators set its goal and initial hypothesis, control deployment and safety, and provide terminal outcomes; evidence covers one task and six specimens.The study does not establish population-level operator effects or a causal estimate of failure supervision. Future tests should use randomized evaluations, multiple operators, robot systems, and tasks while separating research-time agency from control-time authority.Source: Discussion · pp. 11–12

Methods

EndoRSI used EndoForge to iteratively develop policies from source data and physical-trial evidence, while human authorization kept robot execution separate from coding-agent activity. Its methods combined offline policy development, failure-informed updates, and benchtop and ex-vivo evaluation.

EndoForge research loop and coding-agent implementation

  • EndoForge cycled through evidence analysis, candidate implementation, human-authorized physical evaluation, and revision; the agent had no live robot interface while the robot was armed.The operator authorized motion, assigned terminal outcomes, and retained emergency-stop control; completed-trial evidence returned to the coding workspace before the next candidate was prepared.Source: EndoForge research loop and… · p. 12

Source datasets and offline policy development

  • Source tracking and navigation datasets each contained 100 episodes, split by episode into 80 training, 10 validation, and 10 held-out test episodes.Image-conditioned references included behavior cloning, Action Chunking with Transformers, Diffusion Policy, and a vision–language–action policy; validation prompted replacement of the agent’s discrete-bin-plus-residual action representation with continuous prediction.Source: Source datasets and offline policy… · pp. 12–13

Reactive policy synthesis and failure-informed updates

  • The reactive policy paired pretrained Qwen3-VL-4B features [32] with a flow-matching head [33], using source skills and derived stop supervision before physical insertion experience.Gen2 updates cloned successful trajectories, whereas Gen3–Gen4 also used failed-trial avoidance targets that zeroed forward commands in terminal or high-bending states; the separate Success-only lineage discarded failures.Source: Reactive policy synthesis and… · p. 13

Robotic platforms and execution

  • Platform adapters translated shared semantic actions into instrument-specific commands, while the runtime issued zero action when no valid action chunk was available or the watchdog expired.Inference ran asynchronously; fixed command-magnitude and forward-slew limits bounded execution, and human emergency-stop control overrode motor output.Source: Robotic platforms and execution · p. 14

Benchtop and deployment-shift evaluation

  • The final Gen4 checkpoint was evaluated without adaptation in three deployment-shift arms of 21 attempts each; the original Gen4 generation experiment supplied the development-condition reference.The shifts were reduced illumination (distal endoscope light off, ambient lighting unchanged), a 7-mm aperture beside a protruding distractor, and ER2-to-ER3 transfer with task and checkpoint retained. ZS-A was halted after 16 consecutive emergency stops; all 16 attempts remained in analysis.Source: Benchtop and deployment-shift evaluation · p. 14

Ex-vivo evaluation

  • Ex-vivo evaluation used six independent tracheal specimens in two cohorts, with seven attempts per condition per specimen and separate cohort analyses.The initial cohort compared Gen4, Gen1, IBVS, and Success-only; replication added a Success-only configuration control and experienced and novice teleoperation references.Source: Ex-vivo evaluation · p. 14

Outcome assessment and statistical analysis

  • Binary outcomes were independently video-reviewed with 100% agreement across 546 scored trials, while specified two-arm comparisons used two-sided Fisher exact tests.Ex-vivo pooled attempt-level comparisons were exploratory and did not provide specimen-level inference; generation comparisons did not isolate the causal effect of failure supervision.Source: Outcome assessment and statistical… · p. 15

Supplementary Information

The supplementary information lists nine notes on the evidence-to-policy lifecycle, robot systems, experiments, and analysis, alongside three figures on robot configurations and ex-vivo trials.

Supplementary Notes

  • The nine supplementary notes cover the evidence-to-policy lifecycle and authority boundary, robot platforms and software, model selection, task analysis, policy synthesis, experimental protocols, decision chronology, and statistical analysis.The notes also address ex-vivo outcome assessment and rating reliability.Source: Supplementary Notes · p. 19

Supplementary Figures

  • The supplementary figures show endoscopes and robot configurations, an ex-vivo evaluation with a successful Gen4 trial, and representative unsuccessful ex-vivo trials.Source: Supplementary Figures · p. 19

Contents continued

The supplementary tables document the system’s safety authority, experimental and implementation resources, and offline evaluation materials.

Supplementary Tables

  • Table 2 covers control flow and safety authority.Source: Table 2
  • Tables 3–5 cover instruments, experiments, controllers, code, source datasets, and experimental units.Source: Table 3, Table 4, Table 5
  • Tables 6–7 cover model selection, held-out performance, and complete held-out offline results.Source: Table 6, Table 7

Supplementary Note 1: Evidence-to-policy lifecycle and authority boundary

EndoRSI separates human-governed physical execution from agent-led policy development: investigators authorize deployment and retain control, while recorded trial evidence supports subsequent analysis and revision.

Research workflow

  • The EndoForge coding agent used Qwen3.7-Max, while the reactive robot policy separately used Qwen3-VL-4B; the workspace held records, code, validation outputs, and completed-trial evidence.These inputs supported dataset audit, candidate implementation, offline comparison, and revisions.Source: Supplementary Note 1:… · p. 21

Authority boundary

  • During armed trials, candidates remained fixed, the coding agent lacked live robot interfaces, and operators retained reset, termination, and emergency-stop authority.Investigators defined the task and experimental constraints and authorized physical deployment; the runtime translated predictions into bounded motor commands.Source: Supplementary Note 1:… · p. 21

Evidence-to-policy lifecycle

  • After each trial, observations, actions, and human-assigned success, ordinary-failure, or emergency-stop outcomes became evidence for subsequent analysis and policy revision.Investigators supplied the initial align-then-advance decomposition, but no insertion motor targets or physical insertion demonstrations.Source: Supplementary Note 1:… · p. 21

Supplementary Note 2: Robot platforms and software architecture · Supplementary Note 3: Offline model selection

The notes describe how heterogeneous endoscopes share a semantic policy representation while retaining platform-specific mechanics, and how offline candidates were revised using validation evidence. The selected design lowered validation MAE for both tracking and navigation.

Supplementary Note 2: Robot platforms and software architecture

  • ER1, ER2, and ER3 shared one semantic policy interface but used distinct mechanical or firmware implementations across medical, 6.0-mm, and 3.9-mm platforms.ER1 collected tracking data; ER2 supported navigation, development, shifts, and ex-vivo experiments; ER3 supported different-instrument transfer. The physical insertion controller received 400×400 RGB frames at 20 Hz, while training inputs were resized to 224×224.Source: Supplementary Note 2: Robot platforms… · p. 21
  • Canonical semantic records and layered deployment let heterogeneous robots share policy training while preserving device-specific actuation mechanics.Dataset conversion aligned images with actions in [m1, m3, m4]; at inference, a host process selected the newest valid action, enforced limits and a watchdog, and passed commands through platform adapters to lower firmware.Source: Supplementary Note 2: Robot platforms… · p. 21

Source datasets and baseline evaluation

  • Each tracking and navigation dataset contained 100 episodes, split by episode into 80 training, 10 validation, and 10 held-out test episodes.Navigation had 16,534 action rows and tracking had 12,710 aligned image–action rows; task-specific normalization reflected their different motor-unit ranges.Source: Source datasets and baseline evaluation · p. 22
  • Held-out reference comparisons covered CNN behaviour cloning, Action Chunking with Transformers, Diffusion Policy, classical IBVS, and an image-conditioned VLA policy using raw motor-unit MAE, RMSE, and sign accuracy.Copy-previous was a diagnostic for temporal autocorrelation, not a policy reference.Source: Source datasets and baseline evaluation · p. 22

Validation-guided candidate revision

  • V1 was rejected before held-out comparison: its validation MAE was 2.654 for tracking and 19.811 for navigation, versus CNN behaviour cloning’s 2.619 and 19.463.V1 concatenated three RGB frames and predicted eight future actions using 11 bins plus a continuous residual; its post hoc held-out tracking RMSE and sign accuracy were worse than CNN, and it showed no navigation-metric advantage.Source: Validation-guided candidate revision · p. 22
  • Validation MAE fell to 2.426 for tracking and 18.791 for navigation with the selected v2+TE design.The design causally ensembled overlapping action chunks; it used current and temporal-difference images, a ResNet-18 encoder, direct ℓ1 regression of an 8 × 3 chunk, and EMA weights with decay 0.999.Source: Validation-guided candidate revision · p. 22
  • Three independent training runs were evaluated after design selection, with validation records organized by episode rather than treating frames as independent replicates.The cited passage reports this evaluation procedure but does not provide the run-level values.Source: Validation-guided candidate revision · p. 22

Supplementary Note 4: Cross-task analysis and ZS-A implementation · Supplementary Note 5: Reactive VLA synthesis and embodied updates

Cross-task analysis rejected a dark-lumen navigation hypothesis and physical ZS-A trials exposed hazardous forward lunges, redirecting development toward an end-to-end VLA policy. Subsequent updates incorporated physical failures under fixed VLM weights, while execution profiles and instructions changed across generations.

Supplementary Note 4: Cross-task analysis and ZS-A implementation

  • Tracking fits gave m3 ≃23.803ey and m4 ≃32.325ex, while the dark-lumen-centroid navigation hypothesis was rejected (R2 = −0.034 and 0.004).The tracking fits used 10,284 training frames and had R2 = 0.340 and 0.370, respectively.Source: Supplementary Note 4: Cross-task… · pp. 22–23
  • ZS-A achieved 0/16 successes and triggered 16 emergency stops; every trial had a single-tick forward increase of at least 10 motor units.The median maximum increase was 16.7 motor units, and the median first-lunge time was 0.15 s.Source: Supplementary Note 4: Cross-task… · p. 23

Supplementary Note 5: Reactive VLA synthesis and embodied updates

  • The pre-frozen VLA candidate combined tracking alignment, low-bending navigation advancement, and conservative stop targets in an effective 70% source-skills, 20% insertion, and 10% stop-proxy mixture.Its Gen1 policy used pretrained Qwen3-VL-4B features and a flow-matching head for eight future actions; none of the expanded English and Chinese instruction records was a target-task insertion demonstration.Source: Supplementary Note 5: Reactive VLA… · p. 23
  • Failure supervision entered the Gen2-to-Gen3 update: 12 trial records supplied 68 cloning and 144 avoidance samples for a 212-sample, 300-step update.Two success-labelled records supplied cloning targets, ten failure- or emergency-stop-labelled records supplied avoidance targets, and nine records supplied no update sample; action loss decreased from 0.2243 to 0.0918 with VLM weights and normalisation fixed.Source: Supplementary Note 5: Reactive VLA… · p. 23
  • The Success-only control sequence was 1/21, 3/21, 3/21, and 3/21, testing whether repeated success-only updating reproduced the failure-informed generation trajectory.The control used successes from its own preceding physical batch, unlike the Gen1-to-Gen2 update from two successful Gen1 trajectories; success-only denotes the evidence-selection rule, not shared lineage.Source: Supplementary Note 5: Reactive VLA… · p. 23
  • The failure evolution target zeroed forward action for t ≥0.65T, when m1,t > 0 and bending exceeded 0.35, or when absolute bending exceeded 30 motor counts.Text-tagged records contributed no training sample and did not change targets; the update froze the VLM and retained normalisation.Source: Supplementary Note 5: Reactive VLA… · pp. 23–24
  • All 21 Gen3-to-Gen4 trajectories contributed to an update with 240 cloning and 160 avoidance samples; action loss decreased from 0.0914 to 0.0643.The update used 300 steps, batch size 4, and learning rate 10−4.Source: Supplementary Note 5: Reactive VLA… · p. 24

Task instructions and execution profiles during development

  • Gen2–Gen4 used navigation-scale bending, forward gain 1.3, bending gains 1.5 and 2.5, biases −0.048 and 0.111, and 10-s smoothing; replication gave learned policies a common instruction.Gen1-to-Gen2 also changed instruction and execution calibration, while later generations retained that execution profile; the two Success-only replication conditions differed only in execution profile.Source: Task instructions and execution… · p. 24

Supplementary Note 6: Benchtop experimental protocol

The benchtop protocol tested autonomous insertion through a phantom aperture under operator-authorized, safety-constrained execution. It also defined success and failure outcomes and evaluated the frozen Gen4 checkpoint under three deployment shifts.

Benchtop experimental protocol

  • The protocol randomized lateral offset and required explicit operator authorization before the policy alone controlled insertion into a 10-mm phantom aperture.Before each attempt, the operator reset the robot, verified the image stream and zero motor command, and loaded the designated checkpoint and profile; the operator could terminate or emergency-stop the trial.Source: Supplementary Note 6: Benchtop… · p. 24
  • Success required autonomous passage through the aperture; failure and unsafe emergency-stop outcomes were both counted as unsuccessful.The configurable 180-s motor-zeroing threshold did not automatically assign an outcome or close the trial; each reported arm had 21 attempts except ZS-A, halted after 16 consecutive emergency stops.Source: Supplementary Note 6: Benchtop… · p. 25
  • The Gen4 checkpoint was tested without adaptation under reduced illumination, a smaller aperture with a distractor, and a different endoscope.The shifts switched off the distal light, changed the aperture from 10 mm to 7 mm while retaining the phantom material and background, or replaced the 6.0-mm ER2 with a 3.9-mm ER3.Source: Supplementary Note 6: Benchtop… · p. 25

Supplementary Note 7: Human–agent decision chronology · Supplementary Note 8: Ex-vivo protocol and outcome assessment · Supplementary Note 9: Statistical analysis and rating reliability

Human-governed policy development used physical evidence to revise policies, while ex-vivo evaluation and statistical procedures defined the scope and reliability of the reported outcomes. Video and live outcome ratings showed strong agreement, and timing annotations were highly reliable across both ex-vivo cohorts.

Supplementary Note 7: Human–agent decision chronology

  • Validation-driven rejection of v1 preceded continuous action chunks and temporal ensembling; investigators supplied align-then-advance, after which the agent tested task relations and prepared policy candidates.The VLA manifest was fixed before ZS-A physical testing. Gen2 used successful trajectories only; failure-derived avoidance targets first entered Gen3 and were retained in Gen4, distinct from instruction rewriting and execution calibration.Source: Supplementary Note 7: Human–agent… · p. 25

Supplementary Note 8: Ex-vivo protocol and outcome assessment

  • Six discarded porcine tracheae were evaluated in two protocols, with specimens—not repeated insertions—as the independent biological units.The initial block cohort used specimens 1–3 and 84 attempts; the interleaved replication cohort used specimens 4–6 and 147 attempts. Pooled Fisher comparisons treated attempts as technical outcomes and were explicitly exploratory.Source: Supplementary Note 8: Ex-vivo protocol… · p. 25

Initial block cohort

  • All 84 initial-cohort protocol attempts entered analysis; biological n = 3, and two pre-experiment emergency-stop checks were excluded, leaving Gen4 at 21/21.Four policy arms completed seven attempts per specimen in blocks; all three emergency stops among 34 unsuccessful protocol trials were analysed as failures. Navigation time measured travel between the tissue-segment endpoints, not raw video duration.Source: Initial block cohort · pp. 25–26

Interleaved replication cohort

  • The replication cohort evaluated seven conditions across three new tracheae, contributing seven attempts per condition per specimen and 147 protocol attempts overall.Conditions included IBVS, Gen1, Gen4, two Success-only runtime configurations, and experienced and novice teleoperation. The configuration-control contrast changed runtime settings for a fixed checkpoint; configured Success-only versus Gen4 changed the checkpoint under a common runtime profile.Source: Interleaved replication cohort · p. 26

Outcome assessment and human feedback

  • Terminal S/F/E outcomes supplied physical-outcome supervision, and the coding agent could access human feedback on failure cases.Reviewers independently assessed shuffled, identity-free videos using the same success criterion; video-level emergency-stop labels were treated as failures, while safety-event counts came from live records. Successful-trial navigation times used the median of available exact-path annotations; optional free text produced no training samples.Source: Outcome assessment and human feedback · p. 26

Ethics and tissue sourcing

  • No live animals were used or killed, and the porcine stomach and tracheae were discarded post-slaughter tissues from the same vendor.The study involved no human tissue, patients, or clinical decisions; teleoperation arms were descriptive references with one operator in each experience category.Source: Ethics and tissue sourcing · p. 27

Statistical analysis

  • Binary proportions used two-sided Wilson 95% CIs and two-arm comparisons used Fisher exact tests; ex-vivo cohorts were analysed separately.Replication pooled Fisher comparisons were exploratory attempt-level analyses, and offline uncertainty used 20,000 paired bootstrap resamples of 10 held-out episodes per task, not independent frames. No a priori power calculation was performed; ZS-A stopped early for safety, and the single-operator teleoperation references did not support population-level inference about experience.Source: Statistical analysis · p. 27

Rating agreement and timing reliability

  • 546/546 exact-path matches had Fleiss’ κ = 1.00, while replication binary outcomes agreed on 147/147 trials across live and three video ratings.Across 546 scored videos, live outcomes and external evaluator 1 also agreed on all videos (Cohen’s κ = 1.00). Timing ICC(A,1) was 0.994 (bootstrap 95% CI, 0.978–0.998) for 48 initial-cohort successes and 0.956 (0.873–0.984) for 76 replication successes; replication ICC(A,k) was 0.985 (0.954–0.995).Source: Rating agreement and timing reliability · p. 27

Supplementary Figures

The supplementary figures document the endoscope and robot configurations used across experiments and show qualitative examples of successful and unsuccessful ex-vivo trials.

Supplementary Figure 1

  • The figures show ER1, ER2, and ER3 hardware configurations, with ER1 supplying tracking data, ER2 navigation data and policy development, and ER3 different-instrument transfer.ER1 used an Olympus GIF-Q260J with 9.9-mm insertion diameter; ER2 and ER3 used Ailishi S50 endoscopes with 6.0-mm and 3.9-mm diameters, respectively.Source: Figure 1 · p. 28

Supplementary Figure 2

  • Six independent porcine tracheae were used across initial and interleaved replication cohorts, and the figure shows one successful zero-shot Gen4 trial on specimen 4.The trial is a qualitative example and does not add trials or biological replicates; displayed frame times are not scored traversal times.Source: Figure 2 · p. 29

Supplementary Figure 3

  • Representative unsuccessful trials show Gen4 on specimen 6 and Gen1 on specimen 2, with neither terminal view establishing successful lumen alignment.The images alone do not establish the failure mechanism or emergency-stop activation, and the examples add no trials or biological replicates.Source: Figure 3 · p. 30

Supplementary Tables · Supplementary References

The supplementary tables document the evidence-to-policy workflow, safety and experimental design, model validation, and outcomes across benchtop and ex-vivo experiments. Validation rejected the v1 action representation, while direct continuous prediction and temporal ensembling reduced validation error.

Supplementary Tables

  • Supplementary Tables 1–4 describe the evidence-to-policy lifecycle, safety authority and control flow, instrument–experiment–controller mapping, and links among collection data, learned policy, and hardware code.The control-flow caption states that the coding agent acts between trials and has no camera or motor interface while the robot is armed.Source: Table 1 · p. 31, Table 2 · p. 32, Table 3 · p. 32
  • Supplementary Tables 5 and 9–10 cover source-dataset auditing, experimental units, source-task evidence disposition, and generation-specific evidence and update recipes.The generation-table caption notes that lineage comparisons are descriptive because evidence pools and starting checkpoints differed.Source: Table 5 · p. 33, Table 9 · p. 35, Table 10 · p. 35
  • The v1 representation was rejected because its validation MAE exceeded the CNN reference on both tasks, while direct continuous prediction and TE reduced validation error.The supplementary tables report the complete held-out image-conditioned comparison and three-run stability; held-out scores followed validation-based candidate selection.Source: Supplementary Tables · p. 33, Table 7 · p. 34, Table 8 · p. 34
  • Supplementary Tables 11–13 report benchtop outcomes and phantom contrasts, alongside the recorded division of scientific decisions between humans and the agent.The captions specify Wilson confidence intervals, Fisher exact tests, limits on causal interpretation of generation comparisons, and retained human deployment and emergency-stop authority.Source: Table 11 · p. 36, Table 12 · p. 36, Table 13 · p. 37
  • Supplementary Tables 14–15 report initial-cohort and pooled ex-vivo outcomes, treating specimens as biological units and attempts as repeated technical measurements.The pooled cohort analyses are separate and descriptive; the initial Gen4 row contains only its 21 protocol attempts.Source: Table 14 · p. 37, Table 15 · p. 38
  • Supplementary Table 16 describes video-duration phenotypes, while Table 18 reports reliability of successful ex-vivo navigation-time annotations.Video duration is influenced by operator termination and is not a time-to-event efficacy analysis; annotation reliability uses ICC(A,1) and ICC(A,k).Source: Table 16 · p. 38, Table 18 · p. 39
  • Supplementary Table 17 assesses binary outcome agreement across live and video ratings, with emergency-stop counts taken only from the relevant live record.The scored set includes phantom and two ex-vivo cohorts, with complete binary ratings from three selected sources for each trial.Source: Table 17 · p. 39
  • Supplementary Table 19 reports transfer outcomes across two ex-vivo cohorts and six porcine tracheae, with replication Success-only conditions differing in runtime execution profile.Each specimen cell contains seven technical repetitions, and specimens are treated as independent biological units.Source: Table 19 · p. 40
Loading 2610.13376v1…