{"id":"c326f693-9ca9-4f63-9455-b500debc624e","arxiv_id":"2411.13942","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A multi-agent reinforcement learning method that discretizes force feedback into ternary values (-1, 0, 1) makes two-robot cooperative grasping and transport more robust to changes in grasping force, object size, and shape.","lead":"This paper trains two robots to jointly grasp and carry an object using only the direction of force changes, coded as -1, 0, or 1, rather than raw force signals or verbal communication. The authors show in simulation that this simplified signal keeps working when grasping force or object shape changes, and they demonstrate a rough version on real robots.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-robot results overstate success: mean errors 22.7-30.9 cm far exceed the paper's own 6 cm success threshold, and no quantitative raw-force baseline is provided.","rationale":"The reader's weakest assumption correctly identifies the sim2real transfer premise. Our independent pass converges on the same load-bearing point, sharpened by the internal inconsistency between the 6 cm success threshold (Section IV-B) and the reported real-robot errors (22.67 and 30.88 cm in Table V). This is not merely a matter of loose language: the abstract's headline claim ('successfully grasp and transport the object to the target position') is contradicted by the paper's own metric. We also considered whether the ternary representation's consistency under geometry change is a more fundamental flaw: Table IV shows a drop from 82% to 51% success when switching from a rectangular prism to a cylinder, and no direct distributional comparison of ternary observations is provided. However, that drop could partly reflect changed object dynamics rather than representation inconsistency, making it less decisive. The real-robot criterion issue is direct, quantitative, and internal. The simulation experiments are a genuine held-out evaluation, and the ternary representation clearly outperforms raw force under force scaling and geometry shifts, so the method has merit. The appropriate disposition is therefore to keep the reader's CONDITIONAL verdict: the paper should be revised to report real-robot success rates under the 6 cm threshold, provide a quantitative raw-force baseline, and temper the 'successful' claim in the abstract.","tokens_in":7496,"tokens_out":8975,"duration_ms":91969,"concrete_test":"Re-score the existing OptiTrack real-robot logs (or run additional trials) with the Section IV-B success definition: classify each episode as success only if final Euclidean distance between object and target is < 6 cm. Report the success rate and a 95% confidence interval for the mean error for both the proposed method and the raw-force baseline (with the baseline evaluated under identical object sizes and termination conditions). If the success rate is 0/5 (or the confidence interval excludes 6 cm), the claim of successful real-world transportation fails by the paper's own metric and the real-robot section must be revised to describe the result as 'approach with large terminal error' rather than 'successful transport.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV-B, the paper defines a successful task as a final Euclidean distance between object and target under 6 cm. In Section V-B, Table V, the real-robot experiments report mean position errors of 22.67 cm (4x4 cm cross-section) and 30.88 cm (3x3 cm), with variances of 65.32 and 57.19 across five trials. These errors are 3.8-5.1 times the success threshold, so by the paper's own metric none of the real-robot trials reached the target. The abstract and Section V nevertheless state that the robots 'successfully grasp and transport the object to the target position.' This is an internal inconsistency: the central claim of robustness to the inherent sim2real gap rests on experiments that failed the stated success criterion. Furthermore, the raw-force baseline is described only qualitatively ('they only hovered around the object without ever grasping it') with no position-error data, so the real-robot section cannot substantiate the comparative robustness of the ternary representation. The five-trial sample size and large variances also leave the reported means statistically fragile. The load-bearing premise—that the sim2real gap remains inside the policy's robustness margin—is therefore unsupported by the presented measurements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses cooperative grasping and transportation by two robots using force feedback as implicit communication. It introduces a ternary force representation obtained by discretizing the change in gripper force between consecutive time steps into {-1, 0, 1}, and trains MAPPO policies with an asymmetric actor-critic in which the actor observes only the ternary force while the critic also sees the pre-discretization delta force. The authors evaluate the method in Isaac Sim against raw-force, no-force, and ternary-only baselines under variations in grasping force and object geometry, and deploy the simulation-trained policy on two HSRs with GelSight sensors to test object-size variations and sim2real transfer. They report that the proposed representation preserves success under force scaling while raw-force policies degrade, and that the real robots can grasp and transport objects, though with larger position errors than in simulation.","tokens_in":7722,"tokens_out":3888,"duration_ms":40116,"significance":"If the results hold, the ternary representation is an elegant and inexpensive way to make force-based implicit communication invariant to sensor scaling and contact-geometry changes, and the asymmetric actor-critic is a sensible way to mitigate the resulting partial observability. The simulation study is well designed: it compares several baselines, uses 100 test episodes, reports five training seeds, and tests held-out force scales and object geometries. The real-robot deployment is a genuine effort to test sim2real transfer. However, the real-robot evidence as reported does not meet the paper's own success criterion, so the strength of the sim2real claim is currently limited.","major_comments":[{"comment":"The real-robot experiments are not successes by the paper's own metric. Section IV-B defines a successful task as a final Euclidean distance between object and target under 6 cm. Table V reports mean final position errors of 22.67 cm and 30.88 cm, i.e., 3.8 and 5.1 times the success threshold, with variances of 65.32 and 57.19 across five trials. The abstract and Section V-B state that robots \"successfully grasp and transport the object to the target position,\" but the quantitative data show the object did not reach the target in the reported trials. This internal inconsistency affects the load-bearing sim2real claim; the authors should report the percentage of trials satisfying the 6 cm criterion, clearly justify any different success metric used for the real-robot experiments, and adjust the claims accordingly.","section":"V-B, Table V"},{"comment":"The comparison with Raw Force on real robots is only qualitative: the authors write that the raw-force-trained robots \"only hovered around the object without ever grasping it,\" with no position-error data, no success counts, and no number of trials. Since the central claim is that the ternary representation is more robust than raw force, the real-robot section should provide the same quantitative outcome measures for the baseline; otherwise the comparative robustness claim rests only on simulation.","section":"V-B, Raw Force baseline"},{"comment":"The definition of the ternary representation does not specify the threshold that separates 0 from ±1. Raw force signals from real sensors are noisy, so without an explicit deadband the claim that the representation is consistent under small variations is not well-defined. The paper should state the threshold used (or the quantization rule) and ideally provide a sensitivity analysis, since this discretization is the core mechanism of the method.","section":"III-B.1"},{"comment":"The five-trial sample size and the large variances make the real-robot means statistically fragile. With variances of 65.32 and 57.19 over five trials, the standard errors are about 3.6 cm and 3.4 cm respectively, and no confidence intervals or per-trial values are provided, so the difference between the two object sizes is not interpretable as evidence of robustness to object-size variation.","section":"V-B, Table V"}],"minor_comments":[{"comment":"The caption contains a typo: \"Asysmetric\" should be \"Asymmetric\".","section":"Fig. 2 caption"},{"comment":"Table III reports force-observation statistics for grasping force scales 1 and 2, but the success-rate table includes scale 0.5; including the 0.5 case would make the consistency argument more complete.","section":"IV-D.2.a, Table III"},{"comment":"The text and Table V describe object cross-sections of 4×4 cm and 3×3 cm, while the Fig. 5 caption says a \"4 cm-sized object\"; please use consistent dimensional notation.","section":"V-A and Fig. 5 caption"},{"comment":"The phrase \"force signals (without randomization)\" is unclear; specify whether sensor noise was disabled or not randomized during training, since this matters for the sim2real discussion.","section":"IV-B"},{"comment":"The abstract and conclusion claim robustness to object size based on the real-robot experiment, but only two sizes and five trials are reported; the generalization claim should be tempered or supported with more conditions.","section":"I and VII"}],"recommendation":"major_revision","confidential_remarks":"The simulation results provide credible evidence for the core representation benefit under force scaling and partial benefit under geometry changes, and the paper is within scope for a robotics venue. The main problem is that the real-robot section overstates its results relative to the paper's own success metric, and the raw-force real-world baseline is only qualitative. These issues are fixable with additional experiments and a more careful presentation, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core idea is neat: discretize delta force into {-1,0,1} to make force observations invariant to grasp-force scaling and geometry. It's not a deep insight, but the simulation evidence suggests it works. Under held-out changes to grasping force (0.5x, 2x), your method keeps 85%/80% success versus 40%/30% for raw force—that's a compelling difference. The geometry shift (rectangular to cylinder) is weaker (51% vs 25%) but still in the right direction. The asymmetric actor-critic is a sensible way to compensate for the information loss, and the force-observation statistics (Table III) support the claimed mechanism. Simulation support is legitimate.\n\nThe real-robot section is where the paper overreaches. You define success as final position error under 6 cm. The real robot means are 22.67 and 30.88 cm, with variance 65 and 57, from five trials each. That's 3.8–5.1 times your own threshold. By your metric, none of those real trials succeeded. Yet the abstract and Section V claim the robots 'successfully grasp and transport the object to the target position.' There is a mismatch. The Discussion does admit a 'significant discrepancy' and 'notable performance gap,' but that doesn't reconcile the claim with the metric. You also don't report a quantitative raw-force baseline for the real robot—only 'they hovered around the object.' With five trials and no baseline numbers, the sim2real robustness claim is weakly supported at best.\n\nThe core method deserves a serious referee. The contribution is incremental, but legitimate: a simple force pre-processing that buys robustness, demonstrated with a fair held-out evaluation. The real-robot claims need to be rewritten: report the raw baseline, increase trial count if feasible, and either relax the success threshold or describe the results as approximate transport rather than 'success.' If the authors correct the reporting and qualify the real-robot claims, the paper is publishable. I'd send it to review with a request for major revision, not desk reject it.\n\nThis paper is for people working on communication-free multi-robot manipulation and sim2real transfer. The ternary representation is a useful trick to know. I'd probably cite it if I worked in that area, but it's not central to my work.\n\nRecommendation: engage with it as a referee, but treat the real-robot results as preliminary and demand a corrected narrative.","headline":"Solid simulation evidence for a simple ternary force trick, but the real-robot claims overstate what the data show.","tokens_in":8274,"tokens_out":3525,"would_cite":false,"duration_ms":33965,"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 ternary force discretization lets two robots keep coordinating despite changes in grip force, object shape, and size.","keywords":["multi-agent reinforcement learning","ternary force representation","force-sensing feedback","cooperative object transportation","asymmetric actor-critic","sim-to-real transfer","implicit communication"],"falsifier":"Record the ternary observations while replaying the trained policy under a much wider set of contact changes than the paper tests, such as grasping force scaled by 0.25 or 4, a deformable object, or a shifted grasp pose; if the mean and variance of the ternary distribution shift as much as the raw force distribution does, the consistency claim is falsified. A quicker check is to artificially flip the ternary signs during execution and see whether transport success collapses, which would show that the coordination depends on sign accuracy rather than magnitude robustness.","tokens_in":7285,"feed_emoji":"🤖","tokens_out":6792,"duration_ms":66453,"temperature":0.7,"pith_summary":"This paper argues that two robots can coordinate grasping and transporting an object using only force feedback, without explicit communication, if the force signal is reduced to a ternary value indicating whether the force between timesteps increased, decreased, or stayed the same. The central claim is that this ternary representation stays consistent when the grasping force, object geometry, or object size changes, whereas raw force signals shift with those variations and destabilize coordination. To recover the magnitude information lost in discretization, the authors train an asymmetric actor-critic: the actor sees only ternary forces at deployment, while the critic sees both the ternary signal and the full delta force during training. If the claim holds, force-only coordination becomes practical in changing contact conditions and can transfer from simulation to real hardware without retraining.","feed_headline":"Three force values let two robots carry objects together","feed_subtitle":"Discretizing force changes into {-1,0,1} keeps coordination stable when grip force, object shape, or size changes.","key_machinery":"The ternary force representation is the central object: it is the sign of the change in gripper force between consecutive timesteps, taking values $-1$, $0$, or $1$ according to whether the force decreased, stayed the same, or increased. It carries the robustness argument by removing dependence on force magnitude and sensor scale. The supporting mechanism is the asymmetric actor-critic, where the actor observes only $T(\\Delta F_i(t))$ during execution but the critic additionally sees $\\Delta F_i(t)$ during training, supplying the magnitude information needed for credit assignment and mitigating the partial observability introduced by discretization.","core_discovery":"The discovery is that discretizing the delta force $\\Delta F_i(t)=F_i(t)-F_i(t-1)$ into $\\{-1,0,1\\}$ by sign preserves the motion-relevant information for cooperative transport while discarding the magnitudes that depend on grip force, contact area, and object geometry. Using this ternary force as the actor's observation, with the critic given both the ternary and the pre-discretization delta force, the trained policies achieve object transport errors close to the raw-force baseline in simulation and succeed on real two-robot transfers under a different sensor model and different object sizes. The authors show quantitatively that raw force observations have noticeably different mean and variance under different grasping forces, while the ternary representation's distribution changes little, which they identify as the mechanism behind the robustness.","pith_inferences":["The sign-based discretization is a generic scale-invariance trick: it could be applied to any tactile or force feature stream to build coordination signals insensitive to sensor calibration, a generalization the paper does not claim.","The framework's robustness should depend mainly on whether the sign of the delta force flips under new conditions; a natural stress test would be to vary control frequency or add sensor noise until signs flip and measure the resulting success-rate cliff.","The two-robot setting is the basic case; if the representation is truly consistent, extending to more robots with pairwise force observations should behave similarly, but that scaling remains untested.","The paper does not vary the discretization levels; testing two-level or four-level representations would reveal the trade-off between information content and robustness that the ternary choice optimizes."],"forward_implications":["A force-only coordination signal can replace explicit communication for cooperative transport, removing failure modes from communication delays and interruptions.","The same simulation-trained policy can be deployed on a physical robot pair without retraining, even though the real sensor has a different scale, noise, and bias than the simulation sensor.","Changes to grasping force within a factor of two and changes from a rectangular to a cylindrical grasping geometry keep success rates substantially above a raw-force baseline.","Real-robot transport remains functional when the object cross-section changes from 4 cm by 4 cm to 3 cm by 3 cm, although final position error grows relative to simulation.","Because only ternary values are used at execution, the method does not need to maintain calibrated absolute force magnitudes across different contact conditions."],"supporting_citations":[{"why":"Shows force-feedback coordination without explicit communication, the premise the paper builds on.","marker":"[1]"},{"why":"Demonstrates cooperative manipulation exploiting only implicit communication, supporting the force-sensing direction.","marker":"[2]"},{"why":"Previous MARL grasping strategy using observation sharing, which the paper replaces with force-only implicit communication.","marker":"[3]"},{"why":"Provides the multi-agent proximal policy optimization algorithm used for policy training.","marker":"[4]"},{"why":"Shows push-and-pull force discrepancies generating implicit communication in MARL transport, a related precedent.","marker":"[13]"},{"why":"Supplies the asymmetric actor-critic method used to let the actor run on ternary observations while the critic sees richer signals.","marker":"[14]"},{"why":"Frames the inherent simulation-to-real gap in collaborative MARL that the real-robot experiments target.","marker":"[15]"}],"fun_headline_variants":["Discretizing force to -1,0,1 stabilizes cooperative transport","Ternary force sign keeps robots coordinated under grip changes","Robots carry objects robustly with ternary force feedback","Multi-agent RL uses only force sign for robust transport","Force sign beats raw force for robot teamwork"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the distribution of ternary force observations seen during simulation training matches the real world closely enough that a simulation-trained policy acts within its learned distribution on physical robots despite different sensors, friction, and control frequency.","fun_headline_variants_meta":{"raw":{"variants":["Discretizing force to -1,0,1 stabilizes cooperative transport","Ternary force sign keeps robots coordinated under grip changes","Robots carry objects robustly with ternary force feedback","Multi-agent RL uses only force sign for robust transport","Force sign beats raw force for robot teamwork"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000523,"raw_usage":{"total_tokens":2472,"prompt_tokens":832,"completion_tokens":1640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":1562}},"tokens_in":448,"tokens_out":1640,"duration_ms":11837,"temperature":1.0,"reasoning_tokens":1562,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:42:57.984550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the ternary observations while replaying the trained policy under a much wider set of contact changes than the paper tests, such as grasping force scaled by 0.25 or 4, a deformable object, or a shifted grasp pose; if the mean and variance of the ternary distribution shift as much as the raw force distribution does, the consistency claim is falsified. A quicker check is to artificially flip the ternary signs during execution and see whether transport success collapses, which would show that the coordination depends on sign accuracy rather than magnitude robustness.","supporting_citations":[{"cited_title":"Kinematic multi-robot manipulation with no communication using force feedback,","cited_arxiv_id":null,"evidence_quote":"Shows force-feedback coordination without explicit communication, the premise the paper builds on."},{"cited_title":"Cooperative manipulation exploiting only implicit communication,","cited_arxiv_id":null,"evidence_quote":"Demonstrates cooperative manipulation exploiting only implicit communication, supporting the force-sensing direction."},{"cited_title":"A collaborative control method of dual-arm robots based on deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Previous MARL grasping strategy using observation sharing, which the paper replaces with force-only implicit communication."},{"cited_title":"The surprising effectiveness of PPO in cooperative multi-agent games,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-agent proximal policy optimization algorithm used for policy training."},{"cited_title":"Decentralized multi-agent reinforcement learning with global state prediction,","cited_arxiv_id":null,"evidence_quote":"Shows push-and-pull force discrepancies generating implicit communication in MARL transport, a related precedent."},{"cited_title":"Asymmetric actor critic for image-based robot learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the asymmetric actor-critic method used to let the actor run on ternary observations while the critic sees richer signals."},{"cited_title":"Towards closing the sim-to-real gap in collaborative multi-robot deep rein- forcement learning,","cited_arxiv_id":null,"evidence_quote":"Frames the inherent simulation-to-real gap in collaborative MARL that the real-robot experiments target."}],"review_version":1}