{"id":"72161351-d456-4e41-9601-e6067d751a12","arxiv_id":"2505.10251","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SRT-H, a hierarchical language-conditioned imitation learning system, completed all 17 clipping and cutting tasks on 8 unseen ex vivo porcine gallbladders with 100% reported success.","lead":"This paper presents a surgical robot system that plans in language and executes low-level motions, tested on 8 ex vivo pig gallbladders. The authors report 100% success on the clip-and-cut step of cholecystectomy, a step toward longer-horizon autonomous surgery.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'fully autonomous without human intervention' claim is contradicted by the protocol: a human operator loads clips, changes tools, and resumes the system, and the reported duration excludes these interventions.","rationale":"The reader's verdict is CONDITIONAL, and the reader's rationale already notes that an operator handles clip loading, tool changes, and system resumption; however, the reader's formal weakest_assumption focuses on the simulated cutting labels rather than on the autonomy claim. I agree that the simulated cutting data are a real transfer risk, but the more load-bearing issue is the direct mismatch between the paper's headline claim and its own protocol: the system cannot complete the 17-task procedure without repeated human intervention, and the reported timing metric excludes those interventions. This issue is not speculative; it is established by the Methods text, and it strikes at the exact wording of the central claim. A concrete audit of the trial logs would settle it. The paper has genuine strengths: a large real-robot dataset, a hierarchical language-conditioned architecture, and transparent ablations, and the result is plausibly a useful step toward step-level autonomy. For that reason the appropriate disposition remains conditional rather than rejection, and my read does not change the reader's verdict. The simulated-cutting concern remains important and could be tested separately, but the autonomy overclaim is the first thing that must be corrected for the central claim to be accepted as stated.","tokens_in":21579,"tokens_out":7823,"duration_ms":83221,"concrete_test":"Audit the recorded logs for all eight core-experiment trials and tabulate: (1) every operator action (clip loading, tool change, dVRK reinitialization, GUI resume) with timestamps; (2) total wall-clock time including these human interventions, not only robot-motion time; and (3) which of the 17 tasks were executed within a single uninterrupted autonomous segment without an operator step. If any human action appears during the procedure, the 'fully autonomous without human intervention' claim is not supported and should be replaced by a qualified statement such as 'autonomous execution of the 17 robot-motion tasks between operator-performed clip reloads and tool changes.' If the logs are not available, the paper should explicitly state this limitation and refrain from claiming full autonomy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract and Movie 1 is that SRT-H 'achieves a 100% success rate across eight unseen ex vivo gallbladders, operating fully autonomously without human intervention' and completes all 17 tasks. The experimental protocol described in Materials and Methods and Results does not support that claim. An assistant is 'on standby near the dVRK loads another clip and also performs tool changes between clip applier and scissors (filling the role of a surgical nurse)'; the system 'automatically pauses on its own and waits for the operator to load another clip'; and 'the robot autonomously requests for a tool change, and the operator resumes the procedure after making the tool change.' The paper also reports that the dVRK 'had to be reinitialized three times during manual tool changes.' Thus every trial required repeated human interventions, and the reported average of 317 seconds 'excludes the time of reloading the clips and making tool changes performed by the operator.' This is an internal inconsistency between the headline and the protocol, not a matter of consensus: 'fully autonomous without human intervention' is false as stated, and the claimed LoA IV classification is not supported by a protocol in which a human must load every clip, change tools, and resume the system. The underlying learned-policy contribution may still be valuable, but the central claim as written must be revised or the protocol must include autonomous clip loading and tool changing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SRT-H, a hierarchical framework for robotic surgery in which a high-level policy issues language instructions (task-level and corrective) and a low-level policy generates Cartesian-space trajectories from RGB images and wrist-camera views. The framework is evaluated on the clip-and-cut phase of ex vivo cholecystectomy: 16,000 demonstration trajectories from 34 porcine gallbladders are used to train the policies, and eight unseen gallbladders are used for the core evaluation. The authors report a 100% success rate across all 17 tasks with an average duration of 317 seconds, along with ablations on the hierarchical design, wrist cameras, HL DAgger fine-tuning, and an end-to-end baseline, plus a comparison with an expert surgeon.","tokens_in":21863,"tokens_out":4369,"duration_ms":45255,"significance":"If the technical claims can be supported after revision, this is a meaningful advance: it is one of the first demonstrations of a single learned policy executing a long-horizon (17-task) surgical step sequence with self-correction in a realistic ex vivo setting, rather than a single tabletop subtask. The dataset and code are made available, and the ablations provide useful evidence that the hierarchical design and language-based corrective interface contribute to robustness. The comparison with GPT-4o as a planner and the surgeon comparison are useful additional datapoints. The main contribution is the engineering and empirical demonstration of a hierarchical vision-language-action policy for a multi-step surgical procedure, which is timely and relevant to the community.","major_comments":[{"comment":"The central claim \"fully autonomously without human intervention\" is not supported by the protocol described in the manuscript. The operator loads each clip, changes tools between the clip applier and scissors, and resumes the system after tool changes; the dVRK had to be reinitialized three times during manual tool changes; and the reported 317-second duration explicitly \"excludes the time of reloading the clips and making tool changes performed by the operator.\" This is an internal inconsistency, not a matter of interpretation. The abstract and Movie 1 should be revised to state that the system autonomously executes the learned manipulation and strategy-selection steps, with human assistance for clip reloading and tool changes; the LoA IV classification should also be re-examined, since physical assistance with tool management is not merely supervision.","section":"Abstract; Core experiment results; Materials and Methods, Experiment design; Low-level policy, Inference"},{"comment":"The 100% success rate is conditional on excluding approximately 10% of gallbladders with atypical anatomy (artery crossing over the duct or branching). The claim \"generalizing across anatomies\" in the introduction and Movie 1 is therefore limited to the subset of anatomies seen in the training and evaluation sets. The paper should state this exclusion explicitly wherever the 100% rate and the generalizability claims appear, and should temper the clinical-generalization framing accordingly.","section":"Materials and Methods, Data collection, paragraph 2"},{"comment":"The training data for cutting motions were collected without actually closing the scissors: the motion of placing the scissors was performed, and the kinematics were post-processed to simulate cutting. The policy is therefore never trained on real tool-tissue interaction during cutting. The test-time evidence of successful cuts is anecdotal (\"if the cut is not successful, which was very rare\"), and in the surgeon comparison the robot was stopped before closing its grippers to avoid damaging tissue. To support the 100% success rate of the cutting tasks, the paper should provide a systematic verification of cut success (e.g., post-procedure surgical inspection, leakage testing, or video-based evidence that the duct and artery were cleanly severed), and discuss the sim-to-real gap arising from the simulated cutting labels.","section":"Materials and Methods, Data collection, paragraph 3; Comparison with expert surgeon"},{"comment":"The HL policy's corrective-instruction prediction accuracy is reported as approximately 70% (F1 = 56.37%) in Table S2, which is notably lower than the task-instruction accuracy of 97%. The paper should discuss whether this moderate accuracy is sufficient for reliable recovery behavior, and how the 18-way corrective instruction classifier interacts with the LL policy during the reported self-corrections.","section":"High-level policy ablation studies; Table S2"}],"minor_comments":[{"comment":"\"A further extension fo this work\" contains a typo; it should read \"A further extension of this work.\"","section":"Discussion, Limitations, Making SRT-H safer"},{"comment":"The statement that \"the robot autonomously requests for a tool change\" is implemented by a logic-based state machine rather than by the learned policy; this distinction should be clarified so that the autonomy claim is attributed precisely to the learned components.","section":"Low-level policy, Inference"},{"comment":"Table 1 lists 'Duration (s)' but the text clarifies this excludes clip reloading and tool changes; the caption should state this caveat explicitly to avoid misreading.","section":"Table 1 caption and Core experiment results"},{"comment":"Figure 2 is dense; labeling the HL output heads (task, correction flag, corrective instruction) and the selection switch in the panel would improve readability.","section":"Figure 2 caption"},{"comment":"The sentence referring to corrective instructions as \"out of 18 possible motion classes (see Supplementary Methods\" is missing a closing parenthesis and should reference the list of corrective instructions in the supplementary text for completeness.","section":"High-level policy ablation studies"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting, but the 'fully autonomous without human intervention' phrasing is substantially stronger than the experimental protocol, and the simulated cutting data raise a load-bearing validity question about the cutting success rate. I recommend that the editor require the authors to align the abstract, Movie 1, and LoA classification with the actual protocol, and to provide a transparent verification of the cutting outcomes before acceptance. The manuscript is otherwise worth further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SRT-H is a legitimate step forward: it is the first system I know of that uses a language-conditioned hierarchical imitation policy to complete the whole clip-and-cut sequence of cholecystectomy on multiple unseen ex vivo tissues, with mid-procedure self-correction. The paper earns credit for scale: 16,000 trajectories, 17 ordered tasks, ablations against task-only, no-wrist, no-HL-DAgger, and end-to-end variants, plus recovery-from-failure tests and a careful HL-policy ablation. That is real evidence, not a toy.\n\nThe soft spot is exactly what the stress-test says. 'Fully autonomously without human intervention' is contradicted by their own protocol: an operator loads every clip, changes tools, resumes after pauses, and the 317 s average excludes reloading and tool changes. The dVRK also required reinitialization three times. That is not a minor wording issue; it is the paper's headline claim and the basis for the LoA IV classification. The underlying learned policy may still be a major advance, but the claim must be revised or the protocol must include autonomous clip loading and tool changing.\n\nA few lesser concerns. About 10% of gallbladders with atypical anatomy were excluded; that is reasonable for a first study but should be stated wherever 'generalization across diverse tissues' appears. n=8 is small and no uncertainty is given. The training procedure used disabled clip latches and simulated cutting motions without closing the scissors; the authors argue real cuts were usually successful because the scissors were sharp and the policy would retry. That is plausible but the frequency of retry and any failed cuts is not reported. All of these are fixable.\n\nThe circularity concern from the desk review does not really land. Evaluation is on unseen gallbladders, HL DAgger labels are human-corrected language, and corrective labels come from motion trends; that is standard training, not fitting to the test set.\n\nBottom line: this deserves a serious referee, and with revised claims, uncertainty quantification, and a fuller account of operator interventions and excluded anatomies, it could be a strong paper. I would not desk-reject it. I would send it to review with a clear request to fix the autonomy claim.","headline":"A genuine engineering advance in long-horizon surgical imitation learning, but the 'fully autonomous without human intervention' claim is contradicted by the paper's own protocol and needs to be fixed before this can be taken at face value.","tokens_in":22435,"tokens_out":2363,"would_cite":true,"duration_ms":24341,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hierarchical language-conditioned policy is claimed to complete the clip-and-cut phase of cholecystectomy autonomously on unseen ex vivo tissue, with 100% success across eight gallbladders.","keywords":["autonomous surgery","cholecystectomy","hierarchical policy","language-conditioned imitation learning","self-correction","dVRK","ex vivo validation","long-horizon manipulation"],"falsifier":"Run SRT-H on a fresh batch of ex vivo gallbladders with fully functional clips and live scissor closure, and count whether all six clips fully occlude their tubes and both tubes are completely transected without leakage; the 100% claim stands only if every trial passes, and the first incomplete occlusion or partial cut would bound the real-world transfer.","tokens_in":21387,"feed_emoji":"🤖","tokens_out":7804,"duration_ms":67953,"temperature":0.7,"pith_summary":"The paper tries to show that a surgical robot can perform a real, multi-minute surgical step on its own, not just a scripted subtask in a controlled lab. Its proposed system, SRT-H, combines a high-level policy that plans in natural language with a low-level policy that turns each instruction into motions, and it is trained end-to-end from RGB video and language labels by imitation. On eight previously unseen ex vivo porcine gallbladders, the system completed all 17 clip-and-cut tasks of cholecystectomy with no human intervention, self-correcting its errors along the way. If this holds, it would be evidence that step-level autonomy—not just task-level automation—is within reach for soft-tissue surgery, and that language is a workable control interface for long-horizon manipulation.","feed_headline":"Autonomous robot clips and cuts gallbladder 8/8 times","feed_subtitle":"A language-planned surgical policy finished all 17 cholecystectomy steps on unseen tissue with zero human interventions.","key_machinery":"The load-bearing mechanism is the two-level SRT-H stack with natural language as the coupling medium. The high-level policy maps a short history of endoscope frames to three discrete outputs—the next task, a correction flag, and a directional corrective instruction—using a Swin-Transformer vision encoder and a Transformer decoder. The low-level policy is a decoder-only Transformer that fuses wrist-camera and endoscope images, encoded by EfficientNet, with a DistilBERT embedding of the incoming instruction through FiLM layers, and emits a chunk of hybrid-relative actions: translations relative to the endoscope tip, rotations relative to the end-effector, and jaw angles for both arms. Language is doing the decomposition work: it splits a long horizon into named phases, gives the high-level policy a compact way to say 'move the right arm to the right' when the low level drifts, and provides the channel through which DAgger-style human corrections are collected and replayed as training data.","core_discovery":"The central claim, stated on the paper's own terms, is that step-level surgical autonomy is achievable with a hierarchical policy trained purely by imitation from RGB images and language labels. The high-level policy watches endoscope video and issues one of 17 task instructions or one of 18 corrective motion instructions; the low-level policy turns whichever instruction it receives, together with endoscope and wrist-camera images, into a 2-second chunk of Cartesian motions for both arms. On eight ex vivo porcine gallbladders that were not used in training, the system completed all 17 clip-and-cut tasks—grabbing the gallbladder, placing six clips, and cutting the duct and artery twice each—without any human intervention, averaging 317 seconds of robot motion and about six self-corrections per case. The paper also reports that removing the corrective-language channel, the wrist cameras, or the high-level DAgger fine-tuning each degrades success, and that a flat end-to-end low-level-only policy succeeds only 33.3% of the time.","pith_inferences":["Inference: A natural next experiment is to replace the trained high-level policy with a fine-tuned vision-language model emitting the same instruction set; if success holds, the low-level policy is the main carrier of skill, and if it collapses, the high-level policy's learned temporal model of the procedure is essential.","Inference: The current 100% is defined over organs that were selected for their typical anatomy; a deployed system would additionally need to handle the roughly 10% of gallbladders excluded here for artery-over-duct or branched-artery anatomy, since those are common enough to appear in practice.","Inference: Because the high-level policy samples every 3 seconds and the low level commits to 20-30 step action chunks, the framework could be adapted to other multi-phase soft-tissue procedures by changing only the instruction vocabulary and demonstration data, with little architectural change.","Inference: The surgeon comparison suggests SRT-H is smoother but slower; a clinically viable system would need to close that time gap or automate the clip loading and tool changes that currently require a human assistant, since the reported 317 seconds excludes those pauses."],"forward_implications":["If the 100% result holds, an RGB-only, language-conditioned imitation learning stack can handle contact-rich bimanual manipulation on deformable tissue, not just tabletop or navigation tasks.","The hierarchy's self-correction channel means the system can recover from its own errors mid-procedure, shifting the safety question from never making mistakes to detecting and correcting mistakes before tissue damage occurs.","The ablations imply that the corrective-instruction vocabulary, wrist-camera views, and high-level DAgger fine-tuning are each load-bearing; dropping any one measurably lowers success.","The data-scaling curve (66.7%, 77.8%, and 100% success at one-third, two-thirds, and full data) implies that collecting demonstrations from more organs and anatomies is a direct route to broader generalization."],"supporting_citations":[{"why":"Supplies the hybrid-relative action representation and the surgical robot transformer design the low-level policy builds on.","marker":"[26]"},{"why":"Provides the language-correction loop and corrective-instruction labeling used to build the HL DAgger dataset.","marker":"[54]"},{"why":"Motivates the DAgger-style iterative correction procedure used to refine the high-level policy.","marker":"[45]"},{"why":"The prior autonomous bowel anastomosis system this paper contrasts on level of autonomy and controlled conditions.","marker":"[46]"},{"why":"Shows wrist-camera views improve generalization in manipulation, motivating the wrist-camera setup.","marker":"[20]"},{"why":"Supplies the action-chunking formulation and execution-horizon choices used by the low-level policy.","marker":"[64]"}],"fun_headline_variants":["Language-planned bot completes surgery on 8/8 gallbladders","Autonomous surgical robot 100% on unseen gallbladders","Robot surgeon does all 17 steps unaided on 8 cases","Language-conditioned AI performs 17-step surgery at 100%","8/8 ex vivo gallbladders completed autonomously"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system's training data never exercises real clip latching or real scissor closure—clips are disabled and cuts are synthesized in post-processing—so the 100% success claim assumes that transferring to genuine clipping and cutting on unseen tissue requires no dynamics the simulated labels did not capture.","fun_headline_variants_meta":{"raw":{"variants":["Language-planned bot completes surgery on 8/8 gallbladders","Autonomous surgical robot 100% on unseen gallbladders","Robot surgeon does all 17 steps unaided on 8 cases","Language-conditioned AI performs 17-step surgery at 100%","8/8 ex vivo gallbladders completed autonomously"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001161,"raw_usage":{"total_tokens":4808,"prompt_tokens":946,"completion_tokens":3862,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":3773}},"tokens_in":562,"tokens_out":3862,"duration_ms":25843,"temperature":1.0,"reasoning_tokens":3773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:13:06.481719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SRT-H on a fresh batch of ex vivo gallbladders with fully functional clips and live scissor closure, and count whether all six clips fully occlude their tubes and both tubes are completely transected without leakage; the 100% claim stands only if every trial passes, and the first incomplete occlusion or partial cut would bound the real-world transfer.","supporting_citations":[{"cited_title":"Zhao, Samuel Schmidgall, Anton Deguet, Marin Kobilarov, Chelsea Finn, and Axel Krieger","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid-relative action representation and the surgical robot transformer design the low-level policy builds on."},{"cited_title":"Zhao, Archit Sharma, Karl Pertsch, Jianlan Luo, Sergey Levine, and Chelsea Finn","cited_arxiv_id":null,"evidence_quote":"Provides the language-correction loop and corrective-instruction labeling used to build the HL DAgger dataset."},{"cited_title":"A reduction of imitation learning and struc- tured prediction to no-regret online learning","cited_arxiv_id":null,"evidence_quote":"Motivates the DAgger-style iterative correction procedure used to refine the high-level policy."},{"cited_title":"Autonomous robotic laparoscopic surgery for intestinal anastomosis","cited_arxiv_id":null,"evidence_quote":"The prior autonomous bowel anastomosis system this paper contrasts on level of autonomy and controlled conditions."},{"cited_title":"Vision-based manipulators need to also see from their hands","cited_arxiv_id":null,"evidence_quote":"Shows wrist-camera views improve generalization in manipulation, motivating the wrist-camera setup."},{"cited_title":"Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn","cited_arxiv_id":null,"evidence_quote":"Supplies the action-chunking formulation and execution-horizon choices used by the low-level policy."}],"review_version":1}