{"id":"c865a5ea-7c77-47b2-beee-6a8662cb6203","arxiv_id":"2606.06041","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"iCEM+TL applies transfer learning to move iCEM parameters from simple to complex manipulation tasks and uses task decomposition for reward redesign, reporting up to 23% success rate gains in simulation and real-robot validation.","lead":"The paper proposes an iCEM+TL framework that transfers parameters from simpler robotic tasks to complex ones like stacking and shelf placement using transfer learning and reward redesign. A smart generalist might read it to understand practical ways to reduce training time for real robot motion planning.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption directly names the transfer-without-negative-transfer condition that must be true for the 23% claim to be explained by the proposed framework. Because the full text does not supply ablations, parameter tables, or similarity metrics that would secure this condition, the UNVERDICTED verdict and low-confidence assessment remain appropriate. No additional internal flaw was located that would warrant moving the verdict in either direction.","tokens_in":1697,"tokens_out":347,"duration_ms":17850,"concrete_test":"Re-run the downstream tasks with an iCEM baseline that uses the same RR but draws parameters from the upstream distribution instead of the transferred values; if success rates drop by more than the reported 23% margin, the TL component is not carrying the headline improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical one: iCEM+TL (with RR) yields up to 23% higher success rates on stacking/sliding/shelf tasks and transfers to a real Franka. For this to hold, the reported gains must be attributable to the transferred iCEM parameters rather than RR, environment differences, or hyper-parameter search. The abstract and the reader's extraction already flag the transfer assumption as weakest; the full text (per the provided source) does not introduce an internal inconsistency, circularity, or parameter-free derivation that would falsify the claim on its own terms. No machine-checked proof or shipped code is claimed, but the argument is not presented as such. The load-bearing condition therefore remains an empirical one that the current evidence level simply leaves open rather than contradicting.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes the iCEM+TL framework, which augments the Sample-efficient Cross-Entropy Method (iCEM) with zero-shot transfer learning by moving key parameters from simpler upstream tasks to more complex downstream robotic manipulation tasks (stacking, sliding, shelf placement). It further incorporates reward redesign (RR) via task decomposition for stacking and shelf placement. Simulation experiments are reported to yield success-rate gains of up to 23 percent; the approach is additionally validated on a physical Franka Emika robot performing a stacking task.","tokens_in":1826,"tokens_out":421,"duration_ms":21006,"significance":"If the reported gains can be shown to arise specifically from the transferred iCEM parameters rather than from reward redesign, environment differences, or hyper-parameter tuning, the work would offer a practical route to sample-efficient low-level planning that reduces the need for task-specific retraining. The real-robot demonstration, if statistically supported, would strengthen the case for deployability.","major_comments":[{"comment":"Abstract: the central claim of 'success rate improvements of up to 23%' is presented without any baseline (plain iCEM, other planners), number of trials, variance, error bars, or statistical test, rendering the magnitude and attribution of the gain impossible to evaluate.","section":"Abstract"},{"comment":"Results / Experiments: the transfer-learning assumption—that iCEM parameters learned on simpler tasks transfer reliably to stacking/sliding/shelf tasks without negative transfer or further adaptation—is load-bearing yet unsupported by ablations that isolate TL from RR or that quantify transfer success versus failure cases.","section":"Results"}],"minor_comments":[{"comment":"Abstract: the expansion of iCEM is given as 'Sample-efficient Cross-Entropy Method' but the original iCEM reference is not cited, leaving readers without the source of the base algorithm.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive review. We address each major comment below and outline the revisions we will make to strengthen the manuscript.","responses":[{"response":"We agree that the abstract is insufficiently detailed for evaluating the reported gains. In the revised version we will expand the abstract to explicitly name the plain iCEM baseline, state the number of trials, and report variance (or error bars) together with any statistical tests used. The results section already contains these quantities; the abstract will be updated to match.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim of 'success rate improvements of up to 23%' is presented without any baseline (plain iCEM, other planners), number of trials, variance, error bars, or statistical test, rendering the magnitude and attribution of the gain impossible to evaluate."},{"response":"We acknowledge that isolating the contribution of transferred iCEM parameters from reward redesign (RR) is necessary to substantiate the transfer-learning claim. The current experiments evaluate the combined iCEM+TL+RR pipeline; we will add explicit ablation studies in the revised manuscript that separately apply TL without RR and RR without TL, and that report transfer success versus failure rates across the downstream tasks.","revision_made":"yes","referee_comment":"[Results] Results / Experiments: the transfer-learning assumption—that iCEM parameters learned on simpler tasks transfer reliably to stacking/sliding/shelf tasks without negative transfer or further adaptation—is load-bearing yet unsupported by ablations that isolate TL from RR or that quantify transfer success versus failure cases."}],"tokens_in":1325,"tokens_out":360,"duration_ms":24600,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main move is to take the existing iCEM evolutionary planner and transfer its key parameters from simpler upstream tasks to more complex downstream ones like stacking, sliding, and shelf placement. They also add reward redesign through task decomposition on some of those. The reported outcome is up to 23% higher success rates in simulation, with a real-robot check on a Franka Emika arm for stacking.\n\nWhat it does reasonably is show a practical way to reuse planning parameters across task difficulty levels without full retraining. The real-robot validation is a clear plus, since many planning papers stay in simulation. That part gives some evidence the approach can leave the lab.\n\nThe soft spots are in the results. The abstract supplies no baselines, no error bars, no trial counts, and no breakdown separating the effect of transferred parameters from the reward redesign. Without those, the 23% figure is hard to interpret. The central assumption—that parameters from easy tasks will guide complex ones without negative transfer—remains untested in the provided details, so the gains could come from other sources. This is an incremental engineering step rather than a new method or derivation.\n\nThe work is aimed at roboticists already using sample-efficient planners for manipulation who want ideas for cross-task reuse. It shows honest application of known tools to concrete tasks. A serious referee should see the full methods and data to check the experiments, even if the claims need tightening.","headline":"This applies transfer learning to iCEM parameters for harder manipulation tasks plus reward redesign, with a real Franka test, but the evidence for the 23% gains is too thin to evaluate properly.","tokens_in":2349,"tokens_out":376,"would_cite":false,"duration_ms":22274,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Transferring iCEM parameters from simple to complex tasks improves robotic manipulation success rates by up to 23%.","keywords":["transfer learning","motion planning","robotic manipulation","iCEM","sample efficiency","zero-shot transfer","task decomposition"],"falsifier":"Running the complex tasks with transferred parameters and finding success rates equal to or lower than the standard iCEM without transfer.","tokens_in":2605,"feed_emoji":"🤖","tokens_out":537,"duration_ms":20904,"temperature":0.7,"pith_summary":"The paper proposes an iCEM+TL framework that uses transfer learning to move key parameters from easier robotic tasks to harder ones such as stacking and shelf placement. It also redesigns rewards by breaking tasks into parts. This leads to higher success in simulation and works on a real robot arm. A reader cares because it promises faster, more efficient planning without retraining everything from scratch for each new task.","feed_headline":"Transfer learning boosts robot task success by 23%","feed_subtitle":"Parameters moved from simple to complex manipulation tasks cut training needs and work on real hardware.","key_machinery":"iCEM+TL framework that transfers key parameters of the Sample-efficient Cross-Entropy Method from upstream to downstream tasks.","core_discovery":"The iCEM+TL framework explicitly leverages transfer learning where key iCEM parameters are transferred from simpler upstream tasks to guide more complex downstream tasks, combined with reward redesign through task decomposition, resulting in success rate improvements of up to 23% in simulation and practical feasibility on a real Franka Emika robot.","pith_inferences":["Similar transfer strategies might apply to other planning algorithms beyond iCEM.","Task decomposition for reward redesign could extend to additional manipulation scenarios.","Zero-shot transfer may scale to even more complex multi-step robotic operations if parameter selection is refined."],"forward_implications":["Success rates in complex tasks like stacking, sliding, and shelf placement increase by up to 23% in simulation.","The approach demonstrates real-world applicability by succeeding in a stacking task on a physical Franka Emika robot.","Training times decrease through reuse of parameters across tasks.","Low-level real-time planning becomes more effective in sophisticated robotic systems."],"fun_headline_variants":["iCEM+TL transfers parameters from simple to complex tasks achieving 23% success gains","Transfer learning guides iCEM+TL to 23% higher success in manipulation planning","Reward redesign combined with transfer learning improves iCEM by 23% in simulation","Franka Emika robot tests show iCEM+TL feasibility with 23% success gains"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Key iCEM parameters from simpler tasks will transfer reliably to improve performance on complex tasks without negative effects or need for much extra tuning.","fun_headline_variants_meta":{"raw":{"variants":["iCEM+TL transfers parameters from simple to complex tasks achieving 23% success gains","Transfer learning guides iCEM+TL to 23% higher success in manipulation planning","Reward redesign combined with transfer learning improves iCEM by 23% in simulation","Franka Emika robot tests show iCEM+TL feasibility with 23% success gains"]},"model":"grok-4.3","cost_usd":0.007565,"raw_usage":{"total_tokens":3438,"prompt_tokens":609,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":75649500,"prompt_tokens_details":{"text_tokens":609,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2738,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":609,"tokens_out":91,"duration_ms":19998,"temperature":1.0,"reasoning_tokens":2738,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T01:35:48.578857+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the complex tasks with transferred parameters and finding success rates equal to or lower than the standard iCEM without transfer.","supporting_citations":[],"review_version":1}