{"id":"c11587d1-1cb1-43e1-9c1f-a29faa1fc6c5","arxiv_id":"2510.01711","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding a robot-state-aware contrastive loss to VLA training improves manipulation success on RoboCasa-Kitchen and real-robot tasks.","lead":"A new auxiliary loss for robot foundation models, RS-CL, makes vision-language-action policies attend to the robot's own body state by using distances between proprioceptive states as soft contrastive labels. When added to the GR00T N1.5 pipeline it raises RoboCasa-Kitchen success from 65.7% to 69.7% and real-robot task success from 45.0% to 58.3%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-distance soft labels in Eq. 4 are scale-dependent and Table 3a fixes β=1.0 for all targets, so the claimed advantage of current-state distance over alternatives may be a hyperparameter-scale artifact.","rationale":"The paper's empirical case is genuinely broad: consistent gains on RoboCasa-Kitchen (30/100/300 demos), a real-robot boost, and from-scratch evaluations across multiple VLM backbones. The method is simple and plausible, and the CKNNA analysis is a reasonable process check. I do not think the paper is fraudulent or that the headline improvements are fabricated; the concern is narrower. The central novelty over prior contrastive regularizers is the soft weighting in Eq. 4, and Table 3a is the only test isolating this component. That test is confounded by distance scale because β is fixed across supervision targets. The effect size of the state weighting over vanilla InfoNCE is small (69.7 vs 67.3), so an uncontrolled hyperparameter scale could plausibly account for it. This does not invalidate the headline empirical gains, which may come from view cutoff and generic contrastive regularization, but it undercuts the specific 'robot state-aware' mechanism that the paper claims. The reader already flagged the Euclidean-distance assumption and the lack of a metric ablation; my concern sharpens this into a concrete β-confounded comparison. A β sweep and normalized-state rerun would settle whether state weighting is actually the operative ingredient. Since the reader's conditional verdict already requires additional verification, I do not propose moving to reject or accept; the condition stands.","tokens_in":19136,"tokens_out":12402,"duration_ms":148183,"concrete_test":"On RoboCasa-Kitchen (300 demos), re-run the Table 3a ablation for each supervision target (current state, next single action, next action sequence with DTW, vanilla InfoNCE) over β ∈ {0.01,0.1,1,10,100}, tuning γ for DTW if needed, and repeat with per-dimension z-scored proprioceptive states. If next-action or vanilla InfoNCE matches or exceeds current-state distance at its optimal β, or if z-scoring changes the ranking by more than the 2.4-point gap, the state-distance mechanism is not established and the central claim should be narrowed to a generic contrastive-plus-augmentation effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The unique component of RS-CL relative to vanilla InfoNCE is the soft weighting w_ij in Eq. 4. This weighting is only meaningful if Euclidean distance on the raw proprioceptive vector is a stable similarity measure. The vector concatenates end-effector position (meters), 6D rotation (unitless), and gripper state, so unnormalized Euclidean distance gives arbitrary relative weights. The only ablation isolating this component (Table 3a) compares current-state distance to next-action targets while holding β=1.0 for every target (Appendix B.3, where β and τ are fixed). If next-action or DTW distances have a different characteristic scale, the soft weights become either nearly uniform or nearly one-hot, so the comparison is not target-vs-target but target-plus-wrong-scale. Therefore Table 3a does not establish that proprioceptive state is the right supervision signal; it may be the target whose distance scale accidentally matches β=1.0. Since the paper's title and central contribution are specifically about state-aware weighting, this is the weakest load-bearing step in the argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RS-CL, an auxiliary contrastive regularization loss for Vision-Language-Action (VLA) models. The method appends a learnable summarization token to the VLM output, produces embedding pairs via a representation-level 'view cutoff' augmentation, and applies a weighted InfoNCE loss in which the pair weight is a softmax over negative Euclidean distances between the samples' proprioceptive states. The total objective adds this loss, with a cosine-decayed weight, to the standard flow-matching action prediction loss. Experiments on RoboCasa-Kitchen (30/100/300 demos), LIBERO, real-robot pick-and-place and close-lid tasks, and from-scratch VLA training across several VLM backbones report consistent improvements over GR00T N1.5 and other baselines, with ablations of the soft-label target and augmentation type.","tokens_in":19542,"tokens_out":6558,"duration_ms":49231,"significance":"If the reported results hold, RS-CL is a simple, lightweight, and broadly applicable plug-in for VLA fine-tuning, with evidence spanning multiple simulation benchmarks, real-robot tasks, and different VLM backbones. The paper's strengths include detailed training/evaluation settings, use of public datasets and checkpoints, task-wise result tables, and real-robot generalization experiments. However, the central ablation isolating the state-aware weighting is vulnerable to a distance-scale/hyperparameter confound, and the lack of repeated-seed statistics limits confidence in the magnitudes of the gains. These issues are addressable and do not undermine the overall empirical direction, but they need correction before the specific claim about proprioceptive-state supervision can be accepted.","major_comments":[{"comment":"The load-bearing component of RS-CL is the soft weight w_ij, defined via Euclidean distance on raw proprioceptive vectors that mix position, 6D rotation, and gripper state. Table 3a compares 'current state distance' against next-action distances while fixing β=1.0 and τ=0.2 for every target. If the characteristic distance scales of the different targets differ, a fixed β makes the weights either nearly uniform or nearly one-hot, so the observed ordering (state > next action ≈ no soft label) may reflect which target's scale happens to match β=1.0 rather than the semantic content. To support the claim that proprioceptive state is the right supervision signal, please normalize the features, tune or sweep β per target, and report the effective weight distribution (e.g., entropy of w_ij) to demonstrate the comparison is scale-fair.","section":"Eq. (4), Table 3a, App. B.3"},{"comment":"All reported results are single training runs. Real-robot evaluations use 24 trials per task (App. C.3), so differences such as 50.0 vs 58.3 in Fig. 5a are within sampling noise. Simulation results use 1200 trials but a single seed. Please provide mean ± std over at least 3 seeds for the main tables (Tables 1, 2, 3) and the real-robot results; several ablation gaps are only 1–2 points, so without variance estimates it is difficult to judge whether the central improvements are robust.","section":"Tables 1–3 and Fig. 5"},{"comment":"In Table 3b, the 'No augmentation' row (65.3) is below the no-regularization baseline (65.7, Table 3a), yet the text only says that similar representation-level cutoff operations give 'limited improvements' and does not discuss this regression. This suggests the view-cutoff augmentation is essential for RS-CL to help, rather than merely beneficial, and it complicates attributing the end-to-end gain to the state-aware weighting versus the augmentation. This should be discussed explicitly.","section":"Table 3b vs Table 3a"}],"minor_comments":[{"comment":"Grammar: 'VLA models have shown its capabilities' should be 'their capabilities.'","section":"Abstract"},{"comment":"Typo: 'an representation-level augmentation' should be 'a representation-level augmentation.'","section":"Section 2.2"},{"comment":"Line 2 is redundant because line 3 overwrites h. Please clarify whether the action decoder's conditioning input is the output from the sequence with or without the appended summarization token, and align the text with Eq. (2).","section":"Algorithm 1"},{"comment":"The row for layer 24 appears to duplicate the numbers of the layer-18 row while changing only the average. Please verify the entries.","section":"Table 4 (App. A.3)"},{"comment":"The text cites 'RoboBrain (Team et al., 2025)' but Fig. 7 and Table 10 refer to 'RoboBrain2.0.' Keep the naming consistent.","section":"Section 3 baselines"},{"comment":"The DTW distance for 'next action sequence' is said to use an additional temperature γ=10.0, but it is not stated how the DTW distance is normalized before exponentiation with β. This is relevant to the scale-fairness concern above.","section":"App. B.3"},{"comment":"The reproducibility statement mentions implementation details but does not state whether code will be released. A clear code-release plan would strengthen the paper.","section":"Reproducibility statement"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern from the reader's report is valid and is the key issue to resolve. The scale confound in Eq. (4)/Table 3a is load-bearing for the paper's central claim that proprioceptive state is the causally important supervision signal. The paper is otherwise scientifically honest, well-scoped, and has substantial empirical coverage. I would support acceptance after a careful revision that (i) makes the soft-label ablation scale-fair, (ii) adds multi-seed statistics, and (iii) addresses the 'No augmentation' regression in Table 3b."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent empirical paper. The contribution is a weighted InfoNCE auxiliary loss that uses Euclidean distance between proprioceptive states as soft labels, plus a view-cutoff augmentation at the representation level. It consistently improves GR00T N1.5 on RoboCasa-Kitchen (+4–5 points average, +11 on pick-and-place), gives a real-world bump, and transfers to from-scratch VLM backbones. That is genuinely useful and well validated for an empirical methods paper.\n\nWhat is new: the state-distance-weighted soft contrastive loss and the view cutoff are not in the cited prior work; the Table 3/D.1 comparison to plain InfoNCE shows the weighting is adding something, at least in this setup. I believe the main result.\n\nSoft spots: (1) No error bars or multiple seeds anywhere. For a method whose whole claim is consistent improvement, that is a real gap; some task-level numbers look noisy (e.g., Turn Off Stove: 10→10, 14→8, 28→28). (2) Table 3a is the load-bearing piece for 'state distance is the right supervision,' but β=1.0 is held fixed while swapping targets. Euclidean distance on state (meters, unitless rotation, gripper) and on next actions likely live on different scales; the stress-test note is correct that the comparison could be target-plus-wrong-scale. A quick robustness check with normalized state features or per-target β tuning would settle it. I do not think this kills the paper — the benchmark gains stand on their own — but it weakens the mechanistic claim. (3) No code release noted, only a reproducibility statement. For a method paper, that is below the current bar.\n\nThe paper is honestly written: the limitations section is clear, and Appendix D.1 admits CL does not help on LIBERO. No circularity — performance is measured on held-out success.\n\nWho this is for: people training or fine-tuning VLAs, particularly anyone using GR00T-family models. A reader looking for a cheap plug-in regularizer will find this useful; a reader looking for a principled explanation should treat the 'why it works' story as provisional.\n\nVerdict: send to peer review. It deserves referees, and the scale-dependence issue is exactly what reviewers should ask them to address.","headline":"A simple, plausible VLA regularization with consistent gains; the main soft spot is a scale-dependent weighting ablation the authors don't test.","tokens_in":19963,"tokens_out":3615,"would_cite":true,"duration_ms":44748,"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":"Adding a robot-state-aware contrastive loss to VLA training realigns VLM embeddings with proprioceptive state and improves manipulation, especially precise pick-and-place.","keywords":["vision-language-action models","contrastive learning","representation regularization","proprioception","robot manipulation","InfoNCE","flow matching","multi-view augmentation"],"falsifier":"Train the exact RS-CL setup but replace the state distance in Eq. 4 with random noise or permuted state vectors, keeping all other hyperparameters and the view-cutoff augmentation. If success rates on the simulated kitchen benchmark stay near the reported RS-CL level, the improvement is not caused by state-aligned supervision; if the gains disappear, the causal role of the state distance is confirmed.","tokens_in":19088,"feed_emoji":"🤖","tokens_out":5828,"duration_ms":47752,"temperature":0.7,"pith_summary":"Vision-language-action (VLA) models inherit rich visual and semantic representations from pretrained vision-language models, but those representations do not know about the robot's own body. This paper tries to close that gap with a lightweight auxiliary loss, Robot State-aware Contrastive Loss (RS-CL), which pulls VLA embeddings toward embeddings from episodes with similar proprioceptive states, using Euclidean distance between robot states as soft labels. On top of a strong baseline, RS-CL lifts success from 65.7% to 69.7% on a 24-task simulated kitchen benchmark and from 45.0% to 58.3% on real-robot pick-and-place and lid-closing tasks. The largest gains appear in pick-and-place, which the authors attribute to more precise positioning during grasping and placing. If the claim holds, proprioceptive supervision is a cheap and general way to make pretrained VLM features genuinely useful for control.","feed_headline":"Robot-state loss lifts manipulation success 13 points","feed_subtitle":"Auxiliary contrastive loss that weights embeddings by proprioceptive distance sharpens precision of robot pick-and-place.","key_machinery":"The load-bearing object is the weighted InfoNCE loss in Eqs. 3-4 of the paper. Each training sample is summarized by a learned token, projected to an embedding, and the loss compares every sample to an augmented version of every other sample. The weight w_ij = exp(-||q_i - q_j||_2 / beta) is a soft label built from the Euclidean distance between proprioceptive state vectors; it tells the loss which pairs are control-relevant without needing discrete class labels. The companion mechanism is view cutoff, a representation-level augmentation that zeroes out the feature slice of a randomly chosen camera view, so contrastive pairs differ by viewpoint rather than by task meaning. Together they shap","core_discovery":"The authors' central claim is that the conditioning representations inside a VLA model can be made control-relevant by regularizing them with a contrastive loss whose pair weights come from the robot's proprioceptive state. Concretely, RS-CL appends a learnable summarization token to the VLM output, projects the summarized embedding, and optimizes a weighted InfoNCE loss: pairs of samples with close robot states are assigned larger weights and pulled together, while distant states are pushed apart. This runs alongside the standard flow-matching action-prediction loss, with a view-cutoff augmentation that masks one camera view's embedding slice to create cheap contrastive pairs. The authors r","pith_inferences":["The paper's own limitations section notes that object poses and contact forces are excluded. A direct extension would be to feed those signals into the same soft-label weighting, turning RS-CL into a general alignment objective for any control-relevant state.","The Euclidean-distance assumption is the soft spot: if two states are numerically close but require very different actions because the scene or instruction differs, RS-CL will pull their embeddings together erroneously. A testable fix is to weight distances by vision or instruction similarity, or to learn the state metric.","Because view cutoff simulates viewpoint variation, RS-CL may also improve robustness to missing or occluded cameras; the real-robot close-lid result, where the wrist view becomes occluded mid-task, is consistent with that, though the paper does not isolate the augmentation's contribution to occlusion robustness.","The cosine schedule anneals the contrastive weight to zero, so representation refinement matters most early in training. A curriculum that later re-enables state alignment could yield further gains."],"forward_implications":["RS-CL can be dropped into an existing VLA training loop with only an added contrastive path, so the reported gains come without a second training stage or curated robotics reasoning datasets.","Because the loss supervises the representation rather than the action output, it transfers across backbone vision-language models; the from-scratch experiments show gains on several different backbones.","Performance gains concentrate in pick-and-place, so robot tasks whose bottleneck is precise positioning benefit most from representation alignment to state.","Improvements appear at every dataset size tested, from 30 to 300 demonstrations, making the method relevant to low-data robot learning.","The view-cutoff augmentation gives a cheap way to build contrastive pairs from multi-view observations, which may carry over to any multi-camera robot setup."],"fun_headline_variants":["Proprioceptive contrastive loss boosts robot pick-and-place by 13 points","Robot-state-aware regularization lifts VLA success to 58.3% on real tasks","Weighted contrastive loss with proprioceptive distances sharpens robot control","RS-CL: teach VLA models to feel robot state, +13 points real-world"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that Euclidean distance between proprioceptive state vectors is a faithful proxy for control-relevant similarity: samples whose robot states are close are treated as near-duplicates regardless of visual scene or instruction, and if that fails, RS-CL pushes unrelated embeddings together.","fun_headline_variants_meta":{"raw":{"variants":["Proprioceptive contrastive loss boosts robot pick-and-place by 13 points","Robot-state-aware regularization lifts VLA success to 58.3% on real tasks","Weighted contrastive loss with proprioceptive distances sharpens robot control","RS-CL: teach VLA models to feel robot state, +13 points real-world"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000819,"raw_usage":{"total_tokens":3411,"prompt_tokens":723,"completion_tokens":2688,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2601}},"tokens_in":467,"tokens_out":2688,"duration_ms":17921,"temperature":1.0,"reasoning_tokens":2601,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T12:48:52.856962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the exact RS-CL setup but replace the state distance in Eq. 4 with random noise or permuted state vectors, keeping all other hyperparameters and the view-cutoff augmentation. If success rates on the simulated kitchen benchmark stay near the reported RS-CL level, the improvement is not caused by state-aligned supervision; if the gains disappear, the causal role of the state distance is confirmed.","supporting_citations":[],"review_version":1}