{"id":"a3f82f92-8e97-48ed-bbc7-c7e53c95490f","arxiv_id":"2506.19842","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A hierarchical leader-follower Gaussian world model improves multi-task bimanual manipulation success rates over prior single-arm-based methods.","lead":"ManiGaussian++ extends an existing single-arm robot-learning method to two arms by adding a hierarchical Gaussian world model that predicts how a stabilizing arm and an acting arm move objects. It reports large success-rate gains on 10 simulated bimanual tasks and 60% average success on 9 real-world tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation average in Table II is miscalculated (92+20+48)/3=53.33, not 60.00; without error bars or full-task ablation, the claimed 20.2% improvement is not reliably supported.","rationale":"The central claim is that ManiGaussian++ significantly outperforms the state of the art. That claim is fundamentally empirical, and its support is the tables and real-world results. The most load-bearing link is therefore the reliability of those numbers. I found a concrete internal inconsistency in the key ablation table (Table II): the reported 60.00% average for the full method is impossible because (92+20+48)/3 = 53.33. This is not a rounding issue; it is a miscalculation that directly inflates the apparent contribution of the hierarchical world model. Table I has a similar, smaller discrepancy (35.2 vs 35.6), and the real-world average is reported as both 60% and 62.22%. While a single corrected number might not invalidate the method, it shows the numerical reporting cannot be trusted. Without multiple seeds or error bars, the 'significant' improvement over PerAct2 could be noise. The reader's identified weakness (leader-follower causal ordering) is a legitimate modeling concern, but it is less fundamental: even if the causal ordering is imperfect, the method could still work well. The empirical evidence is what should convince us it works, and that evidence is flawed. A concrete, low-cost check is to recompute the averages and, if confirmed, request variance estimates. This concern supports the existing CONDITIONAL verdict, so the verdict remains unchanged.","tokens_in":12132,"tokens_out":7733,"duration_ms":73538,"concrete_test":"Recompute the averages in Table I and Table II from the per-task values; specifically, verify whether Table II row 4 average is 53.33 instead of 60.00. If the correction changes the marginal contribution of the hierarchical world model, rerun the reported experiments (or inspect the released code/results) to obtain per-task success rates from at least 5 seeds with standard deviations; if the 20.2-point advantage over PerAct2 is within one standard deviation, the significance claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.C and Table II present the ablation that attributes the performance gain to the proposed hierarchical Gaussian world model. Row 4 reports success rates of 92, 20, and 48 on sweep-to-dustpan, handover-item, and push-box, with an average of 60.00%. The correct arithmetic mean is (92+20+48)/3 = 53.33%, not 60.00%. This 6.67-point overstatement inflates the marginal contribution of the hierarchical world model (the jump from row 3 to row 4) from a true 26.67-to-33.33 percentage-point difference to a reported 26.67-to-60.00 difference, changing the conclusion. The main results Table I also lists per-task integer success rates for ManiGaussian++ that sum to 352 over 10 tasks, i.e., 35.2%, while the table reports 35.6%. The abstract states a real-world average of 60% while Section IV.E reports 62.22%. These internal inconsistencies mean the numerical basis of the central claim is unreliable. Moreover, the ablation uses a single seed, only 3 of the 10 tasks, with no error bars, and the three tasks were manually selected from subjectively defined categories. Consequently, the evidence that the leader-follower hierarchy is responsible for the gains is not established even if the arithmetic is corrected. The central claim of significant outperformance therefore rests on numbers that are internally inconsistent and statistically unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ManiGaussian++, a bimanual manipulation agent that extends the ManiGaussian framework. The method first builds a task-oriented Gaussian Splatting field from volumetric features, using VLM-generated instance labels to distinguish a stabilizing arm from an acting arm. It then learns a hierarchical Gaussian world model in which a leader deformation model predicts the scene deformation caused by the stabilizing arm, and a follower model predicts the physical consequences of the acting arm conditioned on the leader's output. This future-scene prediction serves as an auxiliary self-supervision for the visual representation used by a PerceiverIO policy. The paper reports a 20.2 percentage-point improvement over PerAct2 on 10 RLBench2 tasks and a 60% success rate across 9 real-world tasks.","tokens_in":12487,"tokens_out":9835,"duration_ms":87619,"significance":"The proposed hierarchy is a reasonable idea: explicitly separating stabilizing and acting arm roles before predicting future Gaussian deformation is more structured than a monolithic world model, and the task-oriented Gaussian Splatting with VLM supervision is a concrete, practical contribution. The paper also provides code, and the experiments cover both simulation and a real dual-arm setup. If the reported results can be reproduced, the method would be a useful step for multi-task bimanual manipulation. However, the quantitative support is weakened by internal numeric inconsistencies in the main tables and by the absence of multiple seeds or error bars, so the central claim of a 20.2% improvement is not currently established at the reported level of reliability.","major_comments":[{"comment":"The reported average for ManiGaussian++ is 35.6%, but the ten per-task success rates (12, 40, 8, 48, 40, 28, 48, 20, 92, 16) sum to 352, an average of 35.2%. The abstract's 20.2% improvement is computed from the reported averages (35.6 versus 15.4), so the central quantitative claim depends on a number that does not match the per-task evidence. Please correct the table and recompute all relative improvements.","section":"Section IV.B, Table I"},{"comment":"Row 4 reports success rates of 92, 20, and 48 with an average of 60.00%, but the arithmetic mean is 53.33%. This overstates the marginal gain of the hierarchical Gaussian world model relative to Row 3: the true gain is 26.67 percentage points (26.67 to 53.33), not the reported 33.33 percentage points (26.67 to 60.00). The ablation also covers only 3 manually selected tasks out of 10, uses a single evaluation seed, and provides no error bars, so even after correcting the arithmetic the attribution of the performance gain to the hierarchical world model is not statistically supported.","section":"Section IV.C, Table II"},{"comment":"The abstract states a 60% average real-world success rate, but Section IV.E reports 62.22% for ManiGaussian++ (and the relative improvements of 100% and 36.57% are computed from 62.22%). The real-robot evaluation uses 10 episodes per task with no per-task breakdown, variance, or multiple seeds. Please reconcile the numbers and report per-task success rates with confidence intervals or at least a per-task table.","section":"Section IV.E and abstract"},{"comment":"The leader-follower decomposition of bimanual dynamics is a strong causal assumption: the follower receives the leader's deformed Gaussian field and both actions, effectively assuming that the stabilizing arm's motion can be predicted first and the acting arm's effect is conditioned on it. The paper provides no experiment that varies this ordering (e.g., predicting acting-arm deformation first, or predicting both simultaneously), so it is unclear whether the reported gains come from this specific inductive bias or from the auxiliary future-prediction objective in general. A comparison against an alternative ordering, or against a non-hierarchical world model with the same Gaussian backbone, is needed to support the mechanistic claim.","section":"Section III.D, Eq. (5)"}],"minor_comments":[{"comment":"The caption says '12 RLBench2 TASK', but the benchmark described in Section IV.A has 10 tasks; please correct the task count.","section":"Table II caption"},{"comment":"The phrase 'without any pertaining on the simulation' should read 'without any pre-training on the simulation'.","section":"Section IV.E"},{"comment":"There are typographical errors: 'Additionaly' should be 'Additionally' and 'underling' should be 'underlining'; a careful proofreading pass would help.","section":"Throughout"},{"comment":"The PerAct2 row appears to have only nine numeric entries as typeset, while the table has ten task columns; please verify that all rows have ten per-task values.","section":"Table I"},{"comment":"The text says predicted future images from multiple views are aligned, but Eq. (8) writes a single-image L2 loss; please clarify whether the loss is summed or averaged over views.","section":"Section III.E, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"The internal arithmetic inconsistencies in Tables I and II are significant and should be resolved before the paper is considered further. Since the code is provided, a straightforward re-evaluation of the reported averages, plus multi-seed runs for the main and ablation results, would be feasible and should be requested. The real-world success-rate discrepancy between the abstract and Section IV.E also needs reconciliation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Guanxing — quick read of ManiGaussian++. The core idea is real: a hierarchical leader-follower Gaussian world model that separates stabilizing and acting arm deformations, with task-oriented splats distilled from VLM labels. That is a sensible inductive bias for bimanual dynamics and a genuine extension of ManiGaussian rather than a repackaging. The writing is clear, the losses are standard, and the code is linked. The ManiGaussian baseline is the obvious comparison, and the paper treats PerAct2 fairly.\n\nBut the numbers as printed do not hold together. Table II reports a 60.00% average for the full model on three tasks; (92+20+48)/3 is 53.33. The 33.33-point boost attributed to the hierarchical world model is actually 26.66 after correction. Table I sums ManiGaussian++ per-task scores to 352, i.e., 35.2%, not 35.6; the abstract's 20.2% improvement over PerAct2 shrinks to 19.8, and the real-world average is stated as 60% in the abstract but 62.22% in Section IV.E. These are not typographical trifles; the headline claim rests on them.\n\nThe deeper problem is statistical. The ablation runs three hand-picked tasks with a single seed and no error bars. Even with corrected arithmetic, we cannot tell if the hierarchical world model is the driver or noise. The real-robot evaluation is 10 episodes per task. For a 20-point claim, seed variance matters. Also, the leader-follower ordering is a strong causal assumption; the paper never tests a reversed or simultaneous ordering, so the mechanism is plausible but unverified.\n\nBottom line: this deserves a serious referee, not a desk reject. The direction is promising and the method is clearly specified. But I would send it back for major revision: fix the tables, rerun ablations with multiple seeds across all ten tasks, and be honest about confidence intervals. A reader in bimanual manipulation or world-model control should engage with the method; just don't quote the headline numbers.\n\nRecommendation: peer review with a request for major revision.","headline":"Clever and clearly specified bimanual extension of ManiGaussian, but the headline numbers contain arithmetic errors and the ablations are too thin to support the 20.2-point gain.","tokens_in":13037,"tokens_out":5411,"would_cite":false,"duration_ms":43176,"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 hierarchical Gaussian world model that predicts future scenes in a leader-follower order improves bimanual robot imitation by 20.2 percent in simulation and reaches 60 percent average success on real-world tasks.","keywords":["bimanual manipulation","Gaussian world model","Gaussian Splatting","imitation learning","future scene prediction","visual representation learning","leader-follower architecture"],"falsifier":"Take a bimanual task with symmetric simultaneous coupling, such as two grippers rotating a single rigid object together where no arm is the clear stabilizer at any point, and train ManiGaussian++ alongside an identical pipeline with a single joint (non-hierarchical) future-prediction model on that task; if the leader-follower version does not beat the joint model in success rate, the causal-ordering assumption is the wrong explanation for the gains reported here.","tokens_in":1681,"feed_emoji":"🤖","tokens_out":2161,"duration_ms":92247,"temperature":0.7,"pith_summary":"This paper claims that the reason bimanual imitation policies fail is that their visual representations do not capture how the two arms interact through the scene over time, and it proposes a fix: a hierarchical Gaussian world model that predicts future scenes in two ordered steps. First a 'leader' predicts how the stabilizing arm deforms a task-oriented 3D Gaussian field, then a 'follower' predicts the physical consequences of the acting arm on top of that deformation. Training the representation with this future-prediction objective, alongside standard behavior cloning, lifts average success from 15.4 percent to 35.6 percent across ten simulated RLBench2 tasks and to a 60 percent average success rate across nine real-robot tasks. The paper's point is that ordered, body-aware future prediction is a representation-level improvement for bimanual manipulation, not just an extra loss term.","feed_headline":"Bimanual robots gain 20 percent from ordered future-scene prediction","feed_subtitle":"Ordered future-scene prediction in a Gaussian world model is the mechanism behind the gain.","key_machinery":"The central object is the hierarchical Gaussian world model built on task-oriented Gaussian Splatting. In this model the scene is a set of Gaussian particles with positions, colors, orientations, scales, opacities, and an instance logit; the instance logit is learned with supervision from a pre-trained open-vocabulary detector prompted with keywords from the human instruction, distinguishing stabilizing arm, acting arm, and target objects. The dynamics are factorized as a leader deformation field $q_{s,\\phi}$, which predicts the $\\mathrm{SE}(3)$ displacement of Gaussians caused by the stabilizing arm, and a follower deformation field $q_{a,\\phi}$, which predicts the acting arm's additional displacement conditioned on the leader's output. Only rigid-body motion of the particles is predicted, with appearance properties held fixed. This factorization is what lets the model separate the question 'what does the stabilizing arm do to the scene?' from 'what does the acting arm do given that?', and it is the inductive bias that the paper argues is missing from prior single-stream world models.","core_discovery":"ManiGaussian++ is an extension of the ManiGaussian framework from single-arm to bimanual manipulation. It builds a task-oriented Gaussian Splatting representation in which each Gaussian particle carries an instance logit distilled from a pretrained vision-language model, labelling which arm or object it belongs to, so the two manipulators and the target can be separated. On top of this, a hierarchical Gaussian world model decomposes the future scene prediction into a leader step and a follower step: the leader deformation field predicts the Gaussian Splatting deformation caused by the stabilizing arm's motion, and the follower deformation field, taking the leader's output as well as both actions, predicts the acting arm's physical consequences. Rendered future images are compared with ground truth through photometric and task-mask losses, forcing the intermediate volumetric representation to encode multi-body spatiotemporal dynamics. The resulting representation is fed to a PerceiverIO policy head that predicts discrete action bins for both arms. The authors' claim is that this two-step, role-aware future-prediction objective is what allows a single policy to generalize across ten simulated and nine real-world bimanual tasks, where prior single-stream world models fail.","pith_inferences":["A testable consequence is that the leader-follower ordering will matter most when the two arms' roles are asymmetric and clear; tasks with symmetric, simultaneous, mutually coupled arm motion may require a bidirectional or joint dynamics model, and on such tasks the reported advantage should shrink or invert.","Since the instance labels come from a pre-trained vision-language model, the method's ceiling is tied to that model's ability to segment the right objects; in clutter or with novel objects the world model may fail to separate bodies, so robustness could be improved by learning the role assignment from the task rather than distilling it.","The world model predicts only rigid-body motion and keeps appearance fixed, so it is likely to underperform on deformable-object or fluid tasks; folding clothes is included in the real tasks, but a stress test on liquids or fabric with visible shape change would clarify the limits of the rigid-Gaussian assumption."],"forward_implications":["Bimanual policies can be improved without new demonstrations: the future-prediction objective provides free supervision from existing multi-view RGB-D data, and the real-world experiments train on only 30 demonstrations per task.","The gains are representation-level rather than policy-architecture-level, so the hierarchical world model can be attached to existing language-conditioned policy heads like PerceiverIO.","Task-oriented instance labels that distinguish stabilizing and acting arms are required for the dynamics factorization to work; ablations show that adding them improves success from 20.0 percent to 26.7 percent on selected tasks, and adding the hierarchy then raises it to 60.0 percent.","Because the model renders future scenes from explicit Gaussian particles, it also yields interpretable future-prediction visualizations that trace which arm causes which scene change, such as the gripper and broom positions in the sweep-to-dustpan task."],"supporting_citations":[{"why":"ManiGaussian, the single-arm Gaussian world model this paper extends; supplies the base framework and the baseline whose bimanual performance drops.","marker":"[30]"},{"why":"PerAct2/RLBench2, the bimanual benchmark, task suite, and main baseline this method must beat.","marker":"[15]"},{"why":"3D Gaussian Splatting, the explicit scene representation whose particles the world model deforms and renders.","marker":"[26]"},{"why":"GroundedSAM, the pretrained open-vocabulary vision-language model used to generate ground-truth instance masks for task-oriented labels.","marker":"[37]"},{"why":"PerceiverIO, the multimodal transformer used as the policy head to select discrete action bins for both arms.","marker":"[24]"}],"fun_headline_variants":["Leader-follower world model lifts bimanual robot success 20%","Hierarchical Gaussian world model powers bimanual robots","Bimanual robots gain from two-step future prediction","ManiGaussian++: ordered world model for dual-arm tasks"],"cache_read_input_tokens":15104,"weakest_assumption_plain":"The load-bearing premise is that bimanual dynamics can be causally ordered: the stabilizing arm's motion deforms the scene first, and the acting arm's effect is generated conditioned on that deformation, so tasks where both arms act simultaneously with mutual coupling would break this inductive bias.","fun_headline_variants_meta":{"raw":{"variants":["Leader-follower world model lifts bimanual robot success 20%","Hierarchical Gaussian world model powers bimanual robots","Bimanual robots gain from two-step future prediction","ManiGaussian++: ordered world model for dual-arm tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2195,"prompt_tokens":1058,"completion_tokens":1137,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":1067}},"tokens_in":674,"tokens_out":1137,"duration_ms":8878,"temperature":1.0,"reasoning_tokens":1067,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:23:45.437488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a bimanual task with symmetric simultaneous coupling, such as two grippers rotating a single rigid object together where no arm is the clear stabilizer at any point, and train ManiGaussian++ alongside an identical pipeline with a single joint (non-hierarchical) future-prediction model on that task; if the leader-follower version does not beat the joint model in success rate, the causal-ordering assumption is the wrong explanation for the gains reported here.","supporting_citations":[{"cited_title":"Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"ManiGaussian, the single-arm Gaussian world model this paper extends; supplies the base framework and the baseline whose bimanual performance drops."},{"cited_title":"Peract2: Benchmarking and learning for robotic bimanual manipulation tasks,","cited_arxiv_id":null,"evidence_quote":"PerAct2/RLBench2, the bimanual benchmark, task suite, and main baseline this method must beat."}],"review_version":2}