{"id":"ccbe7089-3d7b-40ec-be30-34f807d11306","arxiv_id":"2512.19178","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A LoRA-fine-tuned Qwen2.5-VL model generates hierarchical, JSON-formatted robot policies from images and instructions, and updates them online when tasks change (success > 70% in real-world trials).","lead":"This paper fine-tunes a vision-language model to turn language commands and camera images into step-by-step robot plans built from pre-defined actions, and lets the robot re-plan when the command or scene changes mid-task. It reports success rates above 70% across two different robots and everyday tasks such as pick-and-place and handovers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-embodiment success rates do not support generalization: both robot platforms were included in the LoRA fine-tuning data (Sec. III.C.1), so the >68% success on those two platforms (Sec. IV.A.3) may reflect memorization, not transfer to an unseen embodiment.","rationale":"The reader correctly identified the primitive-library limitation as a scope restriction, and that is a real caveat acknowledged by the authors in the Conclusion. However, the primitive-library issue is openly conceded and thus less damaging: the paper does not claim open-set task planning without any library. In contrast, the cross-embodiment generalization claim is presented as a headline result (Abstract, Sec. I, Sec. IV.A.3) without acknowledging that both embodiments were part of the training data. This is a methodological gap that directly undermines a stated contribution. It is also easily testable: a held-out-embodiment experiment would settle it. For these reasons, I consider the cross-embodiment training contamination to be the single most load-bearing concern. My verdict remains CONDITIONAL (unchanged from the reader), because the core idea and the within-embodiment dynamic planning results are still plausible and supported by real-robot trials, but the acceptance conditions should include a proper unseen-embodiment evaluation. I partially agree with the reader because we both flag weaknesses, but we differ on which is most load-bearing: the reader chose primitive sufficiency, while I argue the cross-embodiment protocol is a more serious, unacknowledged flaw.","tokens_in":9508,"tokens_out":5533,"duration_ms":57983,"concrete_test":"Retrain the VLP using LoRA on data from only ONE embodiment (e.g., ANYmal+Z1), with zero samples from HSR. Then run the same three tasks on the HSR platform, using the exact evaluation protocol from Sec. IV.A.3 (25 trials per task). If HSR success is not significantly above the base Qwen2.5-VL baseline (no fine-tuning), the cross-embodiment generalization claim fails and must be retracted. As a control, also evaluate the current model (trained on both) on the same HSR test set to quantify the drop. Report per-platform and per-task success with confidence intervals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim includes 'cross-embodiment generalization' (Abstract, Sec. I). However, the experimental protocol cannot support this claim. The post-training dataset was expanded via 'multi-view acquisition from different robot embodiments' (Sec. III.C.1), which means the fine-tuning data contains images and policies from both ANYmal+Z1 and HSR. The cross-embodiment evaluation (Sec. IV.A.3) then tests the model on those same two platforms. This is at best a test of multi-embodiment adaptability, not generalization to a held-out embodiment. The Related Work section even claims the approach works 'without requiring additional retraining across different embodiments due to the abstraction provided by defined action primitives' (Sec. II), yet the fine-tuning already included both embodiments. If the model had been trained on only one embodiment and then deployed on the other with no retraining, the >68% success would be evidence for the claimed abstraction. As designed, the experiment cannot distinguish between genuine embodiment-invariant planning and the model learning embodiment-specific visual cues or action parameters from the training data. Thus, a major component of the central claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Vision-Language-Policy (VLP) model for dynamic robot task planning. A pre-trained Qwen2.5-VL model is fine-tuned with LoRA on an 800-sample real-world interaction dataset to output hierarchical JSON-formatted policies composed of predefined action and perception primitives. The system is deployed locally and supports dynamic replanning when a new instruction arrives or task state changes. Experiments on two robot platforms (ANYmal+Z1 and Toyota HSR) cover pick-and-place, handover, scene interaction, and dynamic re-planning scenarios. The authors report planning feasibility above 90% and execution success above 70% for fine-tuned models, and success above 68% across embodiments. They also provide comparisons with prior LLM/VLM/VLA approaches.","tokens_in":9792,"tokens_out":4040,"duration_ms":41330,"significance":"If the central claims were fully supported, this would be a useful contribution: it demonstrates that a relatively small VLM (3B), fine-tuned on modest amounts of real-robot data, can produce interpretable, executable policies through a primitive-API abstraction, and can update plans in response to mid-task changes. The real-robot experiments with two platforms, the comparison against the unfine-tuned Qwen2.5-VL baseline, and local deployment are strengths. However, the paper's most distinctive claim — cross-embodiment generalization — is not supported by the current experimental design, because both evaluation platforms appear in the fine-tuning data. The quantitative reporting also contains internal inconsistencies and lacks statistical grounding. The core method remains plausible, but the evidence as presented is not sufficient for the strength of the stated conclusions.","major_comments":[{"comment":"The claim of 'cross-embodiment generalization' is not supported by the experimental protocol. Section III.C.1 states that the post-training dataset was expanded 'via multi-view acquisition from different robot embodiments,' and Section IV.A.3 evaluates the model on exactly the same two platforms (ANYmal+Z1 and HSR). Success rates above 68% on these platforms test multi-embodiment adaptability or memorization of embodiment-specific cues, not generalization to a held-out embodiment. The Related Work statement that fine-tuning enables deployment 'without requiring additional retraining across different embodiments' is therefore not established. To support the claim, the authors should either train on one embodiment and test on the other, include a held-out embodiment, or explicitly revise the claim to 'multi-embodiment adaptability.'","section":"Sec. III.C.1, IV.A.3, Sec. II"},{"comment":"The quantitative summary in the text contradicts Table II. The text states that 'both fine-tuned versions achieve execution success rates above 70%,' but Table II reports VLP(V+D) Pick & Place success of 68%. Similarly, the text states 'planning feasibility rate (exceeding 90%),' but Table II shows several planning rates below 90% (e.g., 88%, 84%). These inconsistencies need to be corrected and the aggregate claims recomputed with per-condition values.","section":"Sec. IV.B.2, Table II"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the success rates. Trial counts are 50 per condition in Table II, 20 in Fig. 8, and 25 in Fig. 9. The word 'significantly' (e.g., 'significantly higher planning feasibility rate') is used without any statistical test. This is especially important because differences between the two fine-tuned variants are small (4–8 percentage points) and could be within chance. Please provide at least bootstrap confidence intervals and, where feasible, a test such as a two-proportion z-test, and define 'planning feasibility' versus 'execution success.'","section":"Sec. IV.B.2, Fig. 8, Fig. 9"},{"comment":"The comparison with prior methods in Fig. 6 does not state the experimental conditions for the baselines. It is unclear whether [11], [22], and [29] were run by the authors with the same primitive library, same scenes, same number of trials, and same prompt formatting, or whether numbers are taken from their papers. Without a detailed protocol, the claim of outperforming existing methods is not quantitatively verifiable. Table I is also purely qualitative checkmarks, which is acceptable as a capability matrix but does not support the 'surpasses state-of-the-art' wording. Please specify the comparison protocol and, if baselines were not reproduced, state this explicitly.","section":"Sec. IV.B.1, Fig. 6"}],"minor_comments":[{"comment":"The Conclusion concedes that the system is 'constrained by the predefined set of action and perception primitives.' This limitation should be stated earlier and more prominently, since it directly qualifes the claim of adaptability to 'novel tasks' and 'open-set' environments in the Introduction.","section":"Sec. V"},{"comment":"The text says '25 trials for each of the three tasks on two different physical robot platforms' but Fig. 9 does not clearly label which tasks correspond to which platform. Please clarify the task-platform mapping and the failure-type definitions.","section":"Sec. IV.A.3, Fig. 9"},{"comment":"'Millisecond-level latency' is mentioned but no timing measurements are provided. A short table of policy-generation latency and replanning latency would strengthen the local-deployment claim.","section":"Sec. IV.A"},{"comment":"The checkmark table would benefit from a legend explaining what each capability means and from citations for 'BT-Planner' and 'LLM-BT,' which are not explicitly identified in the reference list as the source of those rows.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains a useful real-robot study and the fine-tuning approach is plausible. However, the central cross-embodiment generalization claim is not supported by the current protocol, and the quantitative reporting needs correction and statistical grounding. I am not questioning the authors' integrity; the issues are about experimental design and claim strength. I would be willing to reconsider after a revision that either adds a held-out embodiment experiment or substantially moderates the generalization claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a useful systems contribution: it fine-tunes Qwen2.5-VL with LoRA to emit JSON policies over a fixed library of primitives, adds a memory loop for dynamic replanning, and demonstrates real-robot results on two embodiments with reasonable trial counts (50, 20, 25 per condition). The improvements over the unfine-tuned baseline are consistent, and the interpretable hierarchical policy format is a genuine practical advantage over end-to-end VLA models. The local deployment via vLLM is also a plus. All of that deserves credit.\n\nThe soft spots are mostly about what the experiments can actually support. The main one: the paper claims cross-embodiment generalization, but the training data was expanded via multi-view acquisition from both robot embodiments (Sec. III.C.1), and the cross-embodiment test (Sec. IV.A.3) runs on those same two platforms. That is multi-embodiment adaptability, not generalization to an unseen embodiment. The Related Work sentence about working 'without additional retraining across different embodiments' is therefore not supported by the protocol. This is a load-bearing issue for the abstract's claim, though not for the core system story.\n\nOther gaps are more standard: no error bars or significance tests, the Fig. 6 comparison to prior methods doesn't state the baseline methodology, and there's no code/data release. The limitation that the primitive library constrains open-set planning is acknowledged in the Conclusion, which is honest, but it does bound the claims. The self-citations are background and not a problem.\n\nOverall, the central engineering contribution holds up. The paper would benefit from reframing the cross-embodiment claim, or better, retraining on one embodiment and testing on the other. If the authors release the dataset and evaluation code, that would help substantially.\n\nWho's this for: roboticists working on LLM/VLM-based planning, behavior trees, and legged manipulation. A serious referee should see it; with revisions on the generalization claim and a bit more statistical care, it could be a solid venue paper.","headline":"Solid VLP system paper with honest limitations, but the cross-embodiment generalization claim needs rework: both platforms were in the fine-tuning data.","tokens_in":10299,"tokens_out":1719,"would_cite":false,"duration_ms":20162,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single vision-language model, fine-tuned on real-world interaction data, can turn natural-language commands into executable robot policies and revise them when the task changes mid-execution.","keywords":["vision-language model","robot task planning","dynamic replanning","behavior primitives","cross-embodiment generalization","loco-manipulation","policy generation","semantic reasoning"],"falsifier":"Give the deployed system an instruction whose only feasible solution requires an action absent from the primitive library, such as 'screw the cap onto the bottle' where no screwing primitive exists, and observe whether the robot can execute the generated policy. A failure to execute (or a refusal to plan) would show that the system's competence is confined to the library, not generic task planning.","tokens_in":9387,"feed_emoji":"🤖","tokens_out":7765,"duration_ms":70431,"temperature":0.7,"pith_summary":"This paper tries to establish that a vision-language model, post-trained on a modest set of real-world robot interaction samples, can act as a task planner that reads a natural-language instruction and a current camera image, then outputs a structured, executable policy—a sequence of calls to predefined action and perception primitives. The authors claim this design bridges high-level semantic reasoning and low-level robot control without the large datasets and opaque action outputs typical of end-to-end vision-language-action models. They report planning feasibility above 90% and execution success above 70% across pick-and-place, handover, and scene-interaction tasks, and show the model can dynamically re-plan when a new instruction arrives or the task state changes. If correct, the result matters because it suggests interpretable, locally deployable, cross-embodiment task planning can be achieved from hundreds of training samples rather than millions.","feed_headline":"Vision-language model plans and replans robot tasks in real time","feed_subtitle":"Fine-tuned on real interaction data, it reaches over 90% planning feasibility and works across two robot bodies.","key_machinery":"The load-bearing mechanism is the pairing of a fine-tuned vision-language model with a predefined library of behavior primitives—action primitives such as grasp, lift, place, and handover, plus perception primitives for object localization and grasp-point generation. The model's decoder emits a policy as a simplified behavior-tree-like JSON hierarchy executed top-down; each step invokes a primitive, and perception results feed back into the next query. This design is what allows a single model to transfer across embodiments (primitives abstract away platform kinematics) and to re-plan dynamically (the loop re-queries the model on triggers). The paper compares two fine-tuning strategies—apply","core_discovery":"The central claim is that a pre-trained vision-language model, after fine-tuning on real-world robot interaction data (images paired with instructions and manually annotated policies), can generate hierarchical JSON-formatted policies that invoke predefined action and perception primitives to control a robot. The model receives the current instruction, onboard camera image, robot state, and a task memory, and outputs an ordered sequence of behaviors; each behavior is executed by low-level primitives, and feedback from perception and state updates flows back. A 'strategic trigger'—a new instruction or a detected change in task state—pauses execution, refreshes memory, and re-queries the model","pith_inferences":["We infer that the approach's scope is bounded by the primitive library: a task requiring an action the library does not encode (e.g., screwing a lid) would expose a generated but unexecutable policy, so the 'autonomous planning' claim is as strong as the library is complete.","The cross-embodiment result likely reflects the abstraction layer rather than shared low-level control; each new robot still requires a hand-built primitive implementation, so effort shifts from model training to skill engineering.","This design suggests a natural extension: grow the primitive library incrementally and treat the VLP as a router over skills; then the same fine-tuning recipe could extend to long-horizon tasks that the current experiments do not cover.","A conservative reading of the reported numbers is that planning feasibility (the model proposing an executable plan) is the model's contribution, while execution success is co-determined by the perception and control stack; the paper's own failure analysis attributes most failures to perception and execution, not planning."],"forward_implications":["Task planning for everyday manipulation can be learned from a few hundred real-world samples rather than massive robot datasets, because the model only has to choose and order primitives, not generate low-level actions.","A single model can serve multiple robot platforms without retraining, as long as each platform reimplements the same primitive library; the authors report success above 68% on a legged manipulator and a wheeled single-arm robot.","Mid-task changes—a different object to hand over, a new goal, a conditional state to monitor—can be handled by re-querying the model, yielding over 20% higher success than the baseline in dynamic scenarios.","Policies remain interpretable: each step is a named primitive in a JSON hierarchy, so a human can inspect and, in principle, audit the robot's plan before and during execution.","Local deployment on a single GPU achieves millisecond-level policy-generation latency, making real-time replanning feasible without cloud APIs."],"fun_headline_variants":["VLM replans robot tasks mid-execution","Fine-tuned VLM plans, adapts, and replans","Robot planner handles new instructions dynamically","Cross-embodiment policy model re-plans in real time","Vision-language model adjusts robot strategy on the fly"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The predefined library of action and perception primitives is complete enough to express every behavior that the claimed tasks—and any generalization—require; if a task needs an action outside the library, the model can generate a plausible plan that the robot cannot execute.","fun_headline_variants_meta":{"raw":{"variants":["VLM replans robot tasks mid-execution","Fine-tuned VLM plans, adapts, and replans","Robot planner handles new instructions dynamically","Cross-embodiment policy model re-plans in real time","Vision-language model adjusts robot strategy on the fly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1183,"prompt_tokens":722,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":396}},"tokens_in":466,"tokens_out":461,"duration_ms":4768,"temperature":1.0,"reasoning_tokens":396,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:45:34.659010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the deployed system an instruction whose only feasible solution requires an action absent from the primitive library, such as 'screw the cap onto the bottle' where no screwing primitive exists, and observe whether the robot can execute the generated policy. A failure to execute (or a refusal to plan) would show that the system's competence is confined to the library, not generic task planning.","supporting_citations":[],"review_version":1}