{"id":"f06d24ca-999a-43f2-8685-0a8d324152ff","arxiv_id":"2603.08111","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-stage MARL training scheme that separates object-representation learning from coordination learning improves generalization to unseen objects in decentralized multi-robot cooperative transport.","lead":"Scientists teach two robots to carry unfamiliar objects in a warehouse-like task by first learning with 'cheat' information, then learning to infer the object's properties from their own sensors. The result is a three-stage scheme that improves success rates on objects never seen during training, in simulation and on real hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Stage-2 supervision target g_i^t is never defined (Sec. IV-C): the central 'reconstruction of object-dependent representations' mechanism is untestable because the paper does not specify what is reconstructed, what space it lives in, or why a frozen LSTM can infer it from local observations.","rationale":"I read the paper in good faith. It proposes a plausible three-stage CTDE scheme and provides non-trivial empirical support: five-seed simulation with 1000 evaluation trials per seed, failure-mode analysis, and real-robot trials on two unseen objects. The comparative results (Table I) show DeReCo ahead of strong end-to-end baselines, and the real-robot experiment, though small, demonstrates sim-to-real transfer. I do not see an internal inconsistency that invalidates the empirical claims, and I would not reject on the basis of missing code or confidence intervals alone. However, the central mechanism is under-specified. The paper's own description of Stages 2 and 3 depends on a quantity g_i^t that is never defined, so the claim that 'explicitly reconstructing object-dependent representations leads to better generalization' cannot be evaluated. This is precisely the weakest assumption identified by the reader. I considered whether a more load-bearing concern is the sample-efficiency comparison, which omits the cost of Stages 1 and 2 when plotting DeReCo's learning curve; that is also a real weakness, but it does not threaten the generalization claim, and the paper's main headline includes both. The undefined g is more fundamental because it undercuts the causal story. A concrete ablation with a specified g and control targets would settle whether the mechanism as described is responsible for the gains. Therefore I keep the reader's CONDITIONAL verdict rather than moving to ACCEPT or REJECT.","tokens_in":10938,"tokens_out":8262,"duration_ms":78184,"concrete_test":"Fix the Stage-1 network and define g_i^t as the output of the FC branch that encodes the privileged vector p (mass, friction, shape one-hot) before concatenation with the observation branch. Re-run DeReCo under three Stage-2 targets: (A) this g_i^t, (B) the raw p vector, (C) a fixed random vector independent of the object, keeping Stages 1 and 3 identical. If (A) does not significantly beat (C) on average unseen-object success over the same five seeds, the reconstruction mechanism is not the source of the reported gains; if (A) is statistically tied with (B), the method is effectively privileged-parameter regression and should be described as such.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DeReCo's central claim—that decoupling representation learning from coordination learning improves sample efficiency and generalization—rests on Stage 2 teaching an adaptive encoder e_psi to reconstruct an 'object-dependent representation' g_i^t from local observations, and Stage 3 using that reconstruction in place of privileged information. Section IV-C states only that g_i^t is 'computed during rollouts using the encoder in Stage 1'; it never specifies which layer or output of the Stage-1 network produces g_i^t, what information it encodes, or what its dimension/space is. The Stage-1 architecture (Sec. V-A3) mentions an FC 'encoder' but does not define its output as g_i^t. This matters because the method's explanation for its success is that the LSTM recovers object-dependent information (mass, friction, shape-related cues) from the 27-dimensional local observation. If g_i^t is actually an internal feature entangled with o_i^t and the current action, or if it contains privileged quantities not identifiable from local force/pose signals alone (e.g., exact friction coefficient or global object identity), then the MSE objective is not well-posed and the success of Stage 3 cannot be attributed to the claimed decoupling. Without a precise definition, a critic cannot verify soundness or reproduce the method. This is the most load-bearing assumption; if it fails, the central causal mechanism is unsupported, even if the empirical success rates are real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DeReCo proposes a three-stage MARL framework for decentralized two-robot cooperative transport. Stage 1 trains MAPPO-style actors and a centralized critic with privileged object information (mass, friction, shape); Stage 2 trains an LSTM-based adaptive encoder to reconstruct an 'object-dependent representation' g_i^t from each robot's 27-dimensional local observation; Stage 3 fine-tunes the actor with the frozen adaptive encoder while the critic retains privileged information, then deploys actors from local observations only. The paper evaluates the method in Isaac Sim on three training objects and six unseen objects, comparing against five MAPPO baselines, and reports real-robot trials on two unseen objects. The central claim is that this decoupling of representation learning from coordination learning mitigates bidirectional interference and improves sample efficiency and object generalization.","tokens_in":11368,"tokens_out":4382,"duration_ms":41900,"significance":"If the claimed results hold, DeReCo is a practically useful design pattern for object-adaptive decentralized manipulation: it combines privileged-info pretraining, supervised representation distillation, and fine-tuning in a way that is straightforward to implement with standard MAPPO. The paper has several concrete strengths: it reports success rates over 1,000 trials per policy for nine simulated objects, provides a failure-mode analysis, and includes real-robot transfer experiments with two unseen objects. The staged training is not circular: Stage 2 is supervised by targets from the Stage-1 encoder, and success rates are measured independently from the training objective. However, the central mechanism is underspecified: the target quantity g_i^t is never defined, and the quantitative evidence for sample efficiency and for 'outperforming' baselines is incomplete. These issues are load-bearing because they prevent reproduction and because they leave the causal attribution to 'decoupling' unsupported.","major_comments":[{"comment":"The supervised target g_i^t is never defined. The text says only that g_i^t is 'computed during rollouts using the encoder in Stage 1,' but Stage 1's architecture (Section V-A3) is described only as FC layers with 128 units; no layer or output is identified as g_i^t, and no dimensionality, normalization, or semantic content is stated. This matters because Stage 2's MSE objective L(g_i^t, g_hat_i^t) and Stage 3's use of the trained encoder are the core mechanism by which privileged information is supposed to be removed. Without a definition, a reader cannot determine whether g_i^t is recoverable from the 27-dimensional local observation or whether it encodes global/privileged quantities (e.g., exact friction, mass, or object identity) that are not identifiable from local force/pose signals. The authors should specify exactly which encoder output is used as g_i^t, its information content,","section":"Section IV-C, Fig. 2"},{"comment":"All success rates in Table I are point estimates without error bars, confidence intervals, or significance tests, despite being averaged over five independently trained policies and 1,000 trials each. The claim that DeReCo 'outperforms' baselines is not fully supported by the reported numbers: on the seen Board, DeReCo (0.94) is slightly below MAPPO w/o AE (0.95), and on seen-object average DeReCo and MAPPO w PI are tied at 0.91. Several unseen-object differences are also within plausible seed noise. The authors should report standard deviations, bootstrap confidence intervals, or paired significance tests (e.g., across the five seeds) before claiming consistent superiority.","section":"Table I, Section V-C"},{"comment":"The sample-efficiency claim rests on a single reward curve without variance or a quantitative metric. Fig. 4 shows only a mean curve for each method over 50,000 training steps. More importantly, the DeReCo curve is for Stage 3 only; Stages 1 and 2 consume additional environment interactions before Stage 3 begins. Comparing Stage 3's curve to end-to-end baselines that also train for 50,000 steps does not compare total sample budgets. The authors should state the number of environment steps used in Stages 1 and 2, and report a metric such as area under the reward curve or steps to a success-rate threshold computed over the full training pipeline.","section":"Section V-B2, Fig. 4"},{"comment":"Stage 3 is described as 'progressively' removing privileged information, but no progressive schedule, curriculum, or annealing is defined. From the text, it appears that the actor is simply reinitialized from Stage 1 weights and retrained with the frozen adaptive encoder, while the critic continues to use privileged information. If there is no gradual removal, the word 'progressive' is misleading; if there is a schedule, it should be specified. This matters because the claimed decoupling mechanism relies on the manner in which privileged information is removed.","section":"Section IV-D"}],"minor_comments":[{"comment":"The baseline 'MAPPO w/o AE' is used as the strongest comparison in the real-robot experiments, but its architecture is not fully specified relative to DeReCo's actor. Please clarify whether it uses the same local-observation encoder and MLP layers, without the adaptive encoder, so that the comparison isolates the proposed mechanism.","section":"Section V-A2"},{"comment":"PPO/MAPPO hyperparameters (learning rate, discount factor, GAE lambda, clip ratio, entropy coefficient, minibatch size, number of epochs) and the training budget for each stage are not reported. These are needed for reproducibility and for interpreting the sample-efficiency comparison.","section":"Section V-A"},{"comment":"The observation is said to be in R^27, but the listed components sum to 6+3+2+3+7+6 = 27. Please make the decomposition explicit in the text, as some readers may mis-count the previous action or the force channels.","section":"Section V-A1"},{"comment":"In Table II, MAPPO w/o AE on Board reports 0/5 success but a final error of 0.18 +/- 0.06 m; on Frame it reports 0.34 +/- 0.00 m. Please clarify how the error is computed for failed trials, and consider reporting per-trial values or a full success/failure breakdown.","section":"Table II"},{"comment":"The demonstration video is referenced as 'link' with no actual URL. If the video is intended to be part of the submission, provide a working link or remove the reference.","section":"General"},{"comment":"The notation for the mean squared error L(g_i^t, g_hat_i^t) is introduced but the loss is not written as an equation. A short equation, including whether it is averaged over time and over the batch of collected rollouts, would improve precision.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the undefined supervised target g_i^t; this is fixable in a revision by specifying the Stage-1 encoder output used as the target and adding reconstruction diagnostics. The statistical claims also need tightening, but I would not require entirely new experiments. If the authors can supply the definition and revise the quantitative presentation, the paper is potentially acceptable for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies the known teacher-student recipe from Jeon et al. [22] to two-robot cooperative transport: train with privileged object info, distill a representation into an LSTM from local observations, then fine-tune without privilege. The new bit is framing this as decoupling representation learning from coordination learning in MARL, and arguing that MARL non-stationarity makes joint end-to-end training of both harder. That framing is plausible and the experiments are substantial: three seen and six unseen shapes in Isaac Sim, across five seeds, plus real-robot trials on two unseen objects, with failure analysis and a clear sim-to-real result.\n\nThe soft spots are real but proportionate. The load-bearing one is the supervision target g_i^t in Stage 2. Section IV-C says only that it is \"computed during rollouts using the encoder in Stage 1\" and never specifies which layer or output that is, what it encodes, or what space it lives in. Without that, the MSE objective is not well-defined; a reader cannot tell whether the LSTM is learning mass and friction from local signals or something entangled with actions, and the decoupling claim is hard to verify. The stress-test note is right on this. It is not a circularity issue — the experiments measure success independently — but it is a reproducibility and interpretability gap.\n\nThe other issues are secondary. Table I gives no variances despite five seeds; the sample-efficiency claim rests on a single learning curve with no quantitative comparison; and no code or data is released. Novelty is modest relative to [22], though the multi-agent non-stationarity story does add something.\n\nNone of this invalidates the empirical results; the real-robot success (5/5 and 4/5 vs 0/5 for the baseline) is striking. But as written the core mechanism cannot be audited. This is a paper that deserves a serious referee — the problem matters and the experiments are genuine — but the referee should require a precise definition of g_i^t, error bars or per-seed results, and ideally code. If the authors can specify g, the work would be a solid contribution to MARL-based manipulation.\n\nI'd bring it to a reading group if the group does decentralized or manipulation research, but I wouldn't cite it until the g definition is fixed.\n\nRecommendation: send to peer review, expecting major revision.","headline":"A plausible three-stage recipe for object-adaptive multi-robot transport with solid experimental evidence, but the paper never defines the central object-dependent representation g that Stage 2 reconstructs, leaving the core mechanism untestable as written.","tokens_in":11837,"tokens_out":2397,"would_cite":false,"duration_ms":21792,"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":"DeReCo argues that decoupling object representation learning from coordination learning—through a three-stage privileged-training, supervised-reconstruction, and fine-tuning pipeline—lets two mobile manipulators transport objects they have","keywords":["multi-robot cooperative transport","decentralized execution","multi-agent reinforcement learning","object generalization","representation learning","privileged information","adaptive encoder","sim-to-real transfer"],"falsifier":"Train DeReCo with a deliberately broken Stage-2 target (random vectors, or a constant) and check whether Stage-3 performance collapses; alternatively, measure Stage-2 reconstruction error on each unseen object and correlate it per-object with success rate—if episodes succeed despite poor reconstruction, the reconstructed representation is not what carries the generalization.","tokens_in":10851,"feed_emoji":"🤖","tokens_out":6809,"duration_ms":60563,"temperature":0.7,"pith_summary":"This paper claims that the reason robot teams struggle to carry new objects is that today's end-to-end training forces two difficult learning problems—inferring what the object is from partial sensor data, and learning to coordinate with another robot—to be solved simultaneously, so each interferes with the other. Its proposed remedy, DeReCo, runs three stages: first train coordinated behavior while giving the team the object's mass, friction, and shape as privileged information; then train a separate adaptive encoder to reconstruct the object-dependent representation from each robot's own local observations; finally fine-tune the coordination policy using only those reconstructed representations. The reported experiments show higher success rates on three training shapes and six unseen shapes in simulation, and successful zero-shot transfer to two unseen objects on real robots. A reader should care because the recipe, if it holds, is a general way to add object generalization to decentralized manipulation without inventing new learning algorithms.","feed_headline":"Three-stage training lets robot teams carry unseen objects","feed_subtitle":"Separating object recognition from coordination lets robot teams carry new shapes.","key_machinery":"The key mechanism is the adaptive encoder: an LSTM trained by supervised regression to map each robot's 27-dimensional local observation—previous action, arm joint angles, gripper states, offset to the goal, relative object pose, and quantized force readings—to the object-dependent representation g_i^t that Stage 1's network computed from privileged object information. That representation is never explicitly defined; the paper treats it as whatever internal code Stage 1's encoder produces. The adaptive encoder is this code's decoder, and it is the piece that lets the final policy act without privileged information at execution time.","core_discovery":"The central discovery is the empirical claim that decoupling removes 'bidirectional interference': under partial observability, a robot's inferred representation of the object is noisy, and that noise destabilizes coordination learning; at the same time, the non-stationary multi-agent learning dynamics corrupt the representation learner. DeReCo's three-stage training gives each stage a well-posed objective. Stage 1 establishes coordination under reliable object knowledge; Stage 2 solves a supervised regression problem from local observations to the representation Stage 1 used; Stage 3 fine-tunes the policy with the learned encoder and without privileged information. On the paper's measuremen","pith_inferences":["A broader recipe follows if the paper is right: for any MARL task with a latent environment factor, first learn with oracle access to that factor, distill the factor into a local encoder, then fine-tune the policy; the paper's own design is one instantiation of that recipe rather than a special-purpose transport trick.","The paper leaves the supervision target g_i^t undefined, so a fair test of the framework would be to vary what Stage 2 reconstructs (object identity vs. low-dimensional physical parameters vs. the raw Stage-1 code) and watch success rates; this would show which notion of 'object-dependent representation' actually carries the behavior.","A testable extension is to apply the same three-stage scheme to teams larger than two or to objects that change properties mid-episode; the paper's discussion notes that its current MAPPO-based policy input depends on the number of robots."],"forward_implications":["If the decoupling claim holds, end-to-end joint training with domain randomization is not the only route to object generalization in multi-robot transport; a staged privileged-then-reconstruction pipeline can be more sample-efficient.","At execution, each robot needs only its own local observations—no shared state, no communication, no object model or mass/friction measurement.","The generalization result suggests that within the trained ranges of mass and friction, a single policy can transport unseen shapes with useful success rates, not just the exact training objects.","Success on two unseen objects in hardware without any real-world fine-tuning indicates the representation learned in simulation carries over to real sensors.","The failure analysis suggests that remaining errors are mostly transport-precision failures rather than grasp failures, pointing at coordination accuracy as the next bottleneck."],"fun_headline_variants":["Robot teams carry new objects via decoupled learning","Avoiding interference lets robots carry new objects","Decoupling representation and coordination aids robot transport","Separate learning improves multi-robot object transport","Robot teams generalize to new objects via decoupled training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline rests on the assumption that the object-dependent representation g_i^t used as Stage-2 supervision is well-defined, inferable from a robot's 27-dimensional local observation by a frozen LSTM, and sufficient for coordination without privileged information; Section IV-C never defines what g_i^t actually encodes.","fun_headline_variants_meta":{"raw":{"variants":["Robot teams carry new objects via decoupled learning","Avoiding interference lets robots carry new objects","Decoupling representation and coordination aids robot transport","Separate learning improves multi-robot object transport","Robot teams generalize to new objects via decoupled training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"total_tokens":3428,"prompt_tokens":773,"completion_tokens":2655,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2585}},"tokens_in":517,"tokens_out":2655,"duration_ms":18022,"temperature":1.0,"reasoning_tokens":2585,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:33:22.956891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DeReCo with a deliberately broken Stage-2 target (random vectors, or a constant) and check whether Stage-3 performance collapses; alternatively, measure Stage-2 reconstruction error on each unseen object and correlate it per-object with success rate—if episodes succeed despite poor reconstruction, the reconstructed representation is not what carries the generalization.","supporting_citations":[],"review_version":1}