{"id":"93f3eec3-d40e-42d3-a287-cb2169920bed","arxiv_id":"2505.04860","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"D-CODA augments dual-arm eye-in-hand demonstrations with diffusion-synthesized viewpoint-consistent images and constraint-checked action labels, improving policy success rates on most tested simulated and real-world bimanual tasks.","lead":"This paper introduces D-CODA, a diffusion-based method that creates new wrist-camera views for both robot arms along with matching action labels, producing extra training data for bimanual imitation learning. If it works as reported, it could cut the cost of collecting demonstrations for two-armed robots on coordinated tasks like lifting, pushing, and opening drawers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'identical perturbation' coordination claim is not realized by the §4.3 update: sharing one camera-frame ctrans across arms moves the end-effectors differently in world frame, so the key mechanism is unverified and asymmetric tasks are untested.","rationale":"The reader's weakest assumption—that identical perturbations preserve coordination during contact—is the right general area, but the concern is sharper than stated. The paper's own update rule in §4.3 shows that a single camera-frame T cannot preserve the relative gripper pose because the two cameras have different orientations. This is an internally checkable inconsistency between the prose and the math, not just an untested generalization. It weakens the explanatory story for why D-CODA succeeds, although the empirical results still suggest the method is useful on the tested symmetric tasks. The appropriate verdict remains CONDITIONAL: the paper should either clarify that the perturbations are deliberately camera-frame-identical (and explain why this is sufficient), or add an asymmetric coordination task to validate the generalization claim. The lack of code/data and the modest baseline margins in Table 2 reinforce the need for these conditions, but they do not, by themselves, overturn the central empirical finding.","tokens_in":16119,"tokens_out":6422,"duration_ms":71989,"concrete_test":"Using logged or re-simulated contact-rich states from a PerAct2-style task, compute the relative end-effector transform E_r^−1 E_l before and after applying the §4.3 perturbation for each sampled ctrans. If the induced change in relative translation or rotation exceeds the task tolerance (e.g., ball radius or gripper opening) for a material fraction of states, the coordination-preservation mechanism is not realized. Then run an asymmetric bimanual task—one arm stabilizes a tray while the other pushes an object—comparing D-CODA, per-arm Bimanual DMD, and ACT without augmentation under the paper's protocol. If D-CODA does not beat per-arm augmentation there, the method's generality claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism of D-CODA is stated in §4.2 as 'apply identical perturbations to both arms during contact events,' and the paper claims this preserves bimanual coordination. But §4.3 computes perturbed end-effector poses as C·T·(C)^−1·E with the same camera-perturbation transform T applied to both arms. Because the left and right camera poses C_l and C_r differ, T is conjugated by different rigid transforms: the world-frame displacement of each end-effector is R_C · ctrans, not a common world-frame translation. Consequently, the relative pose between the two end-effectors is not preserved; the change depends on the difference between C_l and C_r. Thus the implementation does not actually enforce the 'identical perturbation' that the prose claims, even for the symmetric tasks evaluated. The paper explicitly notes in §2 that it mainly tests tasks where arms move simultaneously, and never evaluates acting-stabilizing asymmetric tasks, despite claiming the approach is 'not task-specific.' If the shared-camera-frame perturbation is the real coordination mechanism, its effect is different from the one claimed, and the generality of the central claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D-CODA, an offline data-augmentation method for eye-in-hand bimanual imitation learning. D-CODA trains a latent diffusion model to synthesize novel, viewpoint-consistent wrist-camera images for both arms, conditioned on camera pose perturbations, while generating corresponding joint-space action labels via a constrained optimization that is designed to preserve bimanual coordination during contact-rich states. The augmented dataset is used together with the original demonstrations to train downstream policies (ACT and π0-FAST). The method is evaluated on five simulated PerAct2 tasks and three real-world tasks, showing improvements over several baselines, with ablations and a generalization study reported in the appendices.","tokens_in":16338,"tokens_out":4768,"duration_ms":46939,"significance":"If the reported results hold, D-CODA offers a scalable and simulator-free data augmentation approach for bimanual manipulation, which is an important practical contribution given the high cost of collecting diverse demonstrations. The paper includes honest limitations, ablations (e.g., the constrained-optimization ablation in Table 3), and a generalization experiment (Appendix D), which strengthen the empirical narrative. However, the central coordination mechanism is misdescribed relative to the implementation, and the evaluation is limited to symmetric tasks, so the generality of the claimed coordination preservation is not established. The method's empirical promise is real, but the paper currently overstates the mechanism by which coordination is enforced.","major_comments":[{"comment":"The key insight stated in §4.2 is to 'apply identical perturbations to both arms during contact events, ensuring coordinated behavior,' but the implementation in §4.3, C·T·(C)^{-1}·E, does not apply identical world-frame perturbations. Since the left and right camera poses C_l and C_r differ, conjugating the same camera-frame T by C_l and C_r produces different world-frame displacements for the two end-effectors, and the relative pose between the end-effectors is not preserved. The augmented contact states may therefore encode coordination patterns that differ from the original demonstrations, contradicting the stated mechanism. The paper should either correct the description to 'identical camera-frame perturbation' and provide direct evidence (e.g., measuring distance and relative orientation between the end-effectors before and after augmentation) that coordination is approximately preserved, or modify the optimization to enforce a genuinely common world-frame translation. As written, the central coordination claim is not supported by the derivation.","section":"§4.2 and §4.3"},{"comment":"The contact detection heuristic based on SAM2 masks, z-score filtering, depth thresholding, and SSIM is a key component of the pipeline because it determines when constrained optimization is applied, yet it is not validated in the paper. There is no evaluation of detection accuracy against ground truth (which is available in simulation) and no ablation showing how misclassification affects downstream performance. Given that random perturbations during contact-rich states are expected to degrade coordination (as the ablation in Table 3 suggests), errors in contact detection could be a significant source of failure. Please report detection accuracy, at least on the simulation tasks, or demonstrate robustness to detection thresholds.","section":"§4.2 (contact detection)"},{"comment":"The paper claims in §2 that the approach is 'not task-specific,' but all five simulation tasks and all three real-world tasks are symmetric in the sense that both arms move simultaneously. The identical-perturbation assumption has not been tested for acting-stabilizing bimanual tasks (e.g., one arm holds an object while the other acts), where a common camera-frame displacement could break the required role asymmetry. Either test at least one asymmetric task or explicitly restrict the claim to symmetric coordination, which would require softening the statement in §2 and the abstract.","section":"§2 and §5"},{"comment":"The empirical claims are based on only three seeds per simulation task and 20 trials per real-world task-condition, and no confidence intervals or significance tests are reported. Several performance gaps are small (e.g., Lift Ball in Table 2: 17/20 vs 15/20), and it is unclear whether these differences are reliable. Reporting binomial confidence intervals or running multiple seeds with variance would substantially strengthen the claim that D-CODA 'outperforms baselines' across tasks.","section":"§5.3 and §5.4"}],"minor_comments":[{"comment":"There is a typo in §3: 'we suppresst and instead use' should be 'we suppress t and instead use'.","section":"§3"},{"comment":"The caption states 'The first black column of images are the original states'; 'column ... are' should be 'column ... is'.","section":"Figure 4 caption"},{"comment":"The notation '∆p = aTb' for a pose transformation is introduced without a clear definition of the matrix product ordering; please clarify whether aTb denotes (T_a)^{-1} T_b or T_a^{-1} T_b, and relate it to the later use of C·T·(C)^{-1}·E.","section":"§4.1"},{"comment":"The constraint 'ctrans∈ [−1, 1]3 and ctrans≥mlb' is ambiguous because mlb is described as a 'magnitude' bound; state clearly whether the inequality is element-wise and whether mlb is a scalar or a vector.","section":"§4.2"},{"comment":"The contact-rich perturbation only uses an identity rotation matrix, while rotation bounds [rlb, rub] are introduced for contactless states. Clarify whether rotation perturbations are ever applied to contact-rich states, and if not, explain why zero rotation is appropriate.","section":"§4.2"},{"comment":"The paper states that ACT is trained on '100 episodes of training data along with their corresponding augmented data,' but does not specify how many augmented episodes are generated per original episode or the total augmented dataset size. This information is needed to assess the data scaling properties of the method.","section":"§5.2"},{"comment":"The paper does not mention whether code is released; the project website is listed, but for reproducibility, please include a link to the code repository in the paper.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a solid empirical study with clear ablations and honest limitations, and the core idea of generating joint-consistent wrist-camera augmentations with action labels is likely to be of interest to the robotics community. However, the mismatch between the stated coordination mechanism and the actual update equation is a substantive correctness issue that should not be glossed over in a journal publication. I recommend major revision so the authors can either redesign the perturbation to actually enforce identical world-frame motion or reframe the method as 'shared camera-frame perturbation with feasibility constraints' and validate that coordination is preserved. The lack of validation for contact detection and the absence of statistical significance reporting are additional points that should be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version first. This paper takes the DMD recipe—diffusion-based novel-view synthesis with action labels—shows how to run it for two wrist cameras at once, and adds a SAM2-based contact detector and a constrained optimization that tries to keep the two arms' perturbations coordinated. The empirical work is solid for the subfield: five simulated and three real tasks, ablations, a generalization experiment, and results that mostly favor D-CODA. The authors also state their limitations honestly. If you work on bimanual imitation learning or offline data augmentation, this is worth reading.\n\nThe genuinely new piece is the handling of contact-rich states. Instead of sampling perturbations independently per arm, they detect contact via segmentation and depth heuristics, then solve a constrained optimization to choose a camera-frame translation applied to both arms. The ablation shows this matters: on the Lift Ball task, dropping the constraint drops success from 73.3% to 57.3%.\n\nNow the soft spots, in rough order of importance.\n\nFirst, the paper's central claim about what the constraint does is not actually what the implementation does. The prose says they 'apply identical perturbations to both arms during contact events' and Appendix G says the distance and orientations of the end-effectors are maintained. But the update is E' = C·T·C^{-1}·E for each arm, with the same camera-frame transform T. Because the two cameras have different poses, a shared T produces different world-frame translations for the two arms, so the relative pose of the end-effectors is not preserved. The magnitudes are small (1–2 cm), so on the symmetric tasks it may not hurt, but the mechanism is not the one described, and there is no measurement showing the relative pose is actually maintained. This should be fixed with either a true world-frame shared translation or a clear statement that the shared-camera-frame perturbation is only a heuristic.\n\nSecond, the contact detector is a heuristic stack—SAM2 masks, z-score filtering, depth thresholds, SSIM fallback—and it is never validated. We don't know how often it fires incorrectly or misses. The ablation shows the constraint helps, but not whether the detector is the right gate.\n\nThird, the paper claims the approach 'is not task-specific' but only tests tasks where both arms move more or less symmetrically. Acting-stabilizing tasks, where one arm holds and the other acts, are explicitly named in the related work but never evaluated. That's a real gap between the claim and the evidence.\n\nFinally, no code or data is released, and the real-world results are 20 trials with no error bars. The comparison with π0-FAST is also oversold: D-CODA ties on Push Block and ties on Lift Drawer if you count failures. The wins are real on Lift Ball, but the summary says 'outperforms baselines on all three tasks.'\n\nNone of this sinks the paper. The core idea is plausible and the results on symmetric tasks are good evidence that D-CODA works in the regime it actually tests. But the coordination mechanism needs a formulation that matches the prose, or at least a validation that the relative pose is preserved, and the task-generality claim should be backed by an asymmetric experiment.\n\nRecommendation: send to review. It's a substantive empirical paper that the community will care about, and the issues are fixable. If it comes to you, push on the mechanism mismatch and ask for code.","headline":"A useful, honest extension of DMD to bimanual eye-in-hand augmentation, but the coordination mechanism is oversold and the central assumption is not validated.","tokens_in":16878,"tokens_out":4766,"would_cite":true,"duration_ms":47212,"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":"Diffusion-based offline augmentation improves bimanual wrist-camera policies by synthesizing coordinated views and feasible joint-space actions.","keywords":["bimanual manipulation","data augmentation","eye-in-hand imitation learning","diffusion models","novel view synthesis","contact-rich manipulation","constrained optimization","action labeling"],"falsifier":"Choose an asymmetric contact-rich task in which the left arm holds an object still while the right arm manipulates it; train D-CODA with the identical-perturbation rule, compare its rollout success against a variant that perturbs only the acting arm, and measure the inter-end-effector distance distribution in augmented contact states. If the identical-perturbation variant does not match or beat the alternative, the paper's coordination constraint fails.","tokens_in":1522,"feed_emoji":"🤖","tokens_out":4520,"duration_ms":113798,"temperature":0.7,"pith_summary":"This paper tries to establish that a diffusion model trained only on existing demonstrations can generate additional training data for two-arm robots by simultaneously synthesizing new left- and right-wrist images and new joint commands. The proposed method, D-CODA, is fully offline: it needs no simulator, no environment resets, and no extra teleoperation. Its key move is to treat contact-rich and contact-free states differently—random camera shifts in free space, identical constraint-optimized shifts during gripper-object contact—so the synthesized images and actions respect bimanual coordination. The authors report higher downstream task success than both baseline augmentations and simply adding more demonstrations across most simulated and all real-world tasks tested. If correct, this is a scalable route to better bimanual policies from small datasets.","feed_headline":"Diffusion augmentation lifts two-arm robot task success","feed_subtitle":"Offline method generates coordinated wrist-camera views and feasible joint actions from existing demos.","key_machinery":"The central machinery is a paired latent diffusion denoiser that, conditioned on source wrist images $I^l_a, I^r_a$ and the relative camera transformations $\\Delta p^l, \\Delta p^r$, denoises the latents of both target wrist images in one pass, so the generated left and right views are coherent views of the same perturbed scene. The second mechanism is contact-aware camera perturbation sampling: a large video segmentation model tracks the grippers to split states into contactless and contact-rich; contactless states sample random translation and rotation perturbations, while contact-rich states solve a constrained optimization over translation $c_{\\mathrm{trans}}$ that penalizes table collisions, end-effector proximity, and kinematically infeasible poses, applying the same perturbation to both arms. An inverse-kinematics solver converts the resulting camera-perturbed end-effector pose into a new joint-space action label, and augmented states replace original states every $k$ timesteps to address compounding errors.","core_discovery":"The paper claims that D-CODA, a diffusion-based offline augmentation method, extends a single-arm viewpoint-augmentation recipe to bimanual manipulation. It synthesizes both left and right wrist-camera images in a single conditioned diffusion pass from source images and relative camera transformations, then derives joint-space action labels by applying the same camera perturbation to both arms and solving inverse kinematics rather than reusing the original actions. Segmentation-based contact detection decides when to use random perturbations (contactless states) versus constrained, identical perturbations (contact-rich states). Trained with an action-chunking transformer policy, the augmented dataset outperforms no-augmentation, more-data, and prior viewpoint-augmentation baselines on 4 of 5 simulated tasks and all 3 real-world tasks, and it also boosts a vision-language-action policy on 2 of 3 real-world tasks.","pith_inferences":["Editorial extension: the identical-perturbation rule is a symmetry assumption; for tasks with asymmetric act-stabilize roles, a more promising variant would perturb the acting arm and freeze or minimally move the stabilizing arm, and the paper's own segmentation pipeline could supply the role assignment.","Editorial extension: the paper's generalization experiment shows that the diffusion model can adapt to a new task with only 10 target demonstrations, suggesting D-CODA could be used as a cross-task augmentation engine trained once on several bimanual tasks and fine-tuned briefly on a new one.","Editorial extension: the real-world failure mode of large policy actions on augmented states suggests that pairing D-CODA with smoother action tokenization could close the gap seen with the vision-language-action policy.","Editorial extension: a testable possibility is iteratively augmenting the augmented dataset–adding synthesized states back into the diffusion model's training data–until the pose distribution drifts too far from valid scenes."],"forward_implications":["D-CODA produces an augmented bimanual dataset entirely offline, so any wrist-camera imitation learning pipeline can be supplemented without additional robot time, simulators, or environmental resets.","The contact-aware identical-perturbation constraint is essential: the ablation without constrained optimization drops simulation success on the ball-lifting task from 73.3 percent to 57.3 percent.","D-CODA also improves tasks that do not require symmetric simultaneous lifting, such as dual button pushing and rope straightening, suggesting the augmentation is not limited to coordinated lifts.","Removing the constrained perturbation sampling or swapping the pretrained image encoder degrades synthesized image quality and downstream success, indicating that both components carry the gains.","The method inherits a distributional caveat: it relies on novel camera poses being similar enough to training poses, so very different viewpoints would likely degrade performance."],"supporting_citations":[{"why":"Supplies the conditional diffusion recipe for single-arm wrist-camera view augmentation that D-CODA extends to two arms.","marker":"[9]"},{"why":"Provides gripper segmentation masks used to detect contact-rich vs. contactless states.","marker":"[14]"},{"why":"Defines the action-chunking transformer policy that serves as the downstream evaluation backbone.","marker":"[44]"},{"why":"Supplies the global optimization algorithm used for constrained camera perturbation sampling.","marker":"[75]"},{"why":"Provides the five simulated bimanual tasks and their definition.","marker":"[76]"},{"why":"Provides the simulation environment and waypoint-based demonstration generation.","marker":"[77]"},{"why":"Provides the view-invariant baseline using fine-tuned novel view synthesis, which D-CODA is compared against.","marker":"[12]"},{"why":"Provides the vision-language-action model used as an alternative downstream policy.","marker":"[80]"}],"fun_headline_variants":["Diffusion generates coordinated wrist views for dual-arm training","D-CODA: Synthetic eye-in-hand images for bimanual imitation","Diffusion-based augmentation boosts bimanual robot task success","Coordinated diffusion augments bimanual demonstrations","D-CODA: Diffusion-crafted demos improve two-arm policies"],"cache_read_input_tokens":19072,"weakest_assumption_plain":"The load-bearing assumption is that applying the same camera perturbation to both arms during any gripper-object contact preserves the coordination encoded in the original demonstration; for tasks in which one arm stabilizes while the other acts, this symmetry may encode contradictory augmented states and the downstream policy could degrade.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion generates coordinated wrist views for dual-arm training","D-CODA: Synthetic eye-in-hand images for bimanual imitation","Diffusion-based augmentation boosts bimanual robot task success","Coordinated diffusion augments bimanual demonstrations","D-CODA: Diffusion-crafted demos improve two-arm policies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001289,"raw_usage":{"total_tokens":5263,"prompt_tokens":941,"completion_tokens":4322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":4239}},"tokens_in":557,"tokens_out":4322,"duration_ms":29669,"temperature":1.0,"reasoning_tokens":4239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:19:25.034933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose an asymmetric contact-rich task in which the left arm holds an object still while the right arm manipulates it; train D-CODA with the identical-perturbation rule, compare its rollout success against a variant that perturbs only the acting arm, and measure the inter-end-effector distance distribution in augmented contact states. If the identical-perturbation variant does not match or beat the alternative, the paper's coordination constraint fails.","supporting_citations":[{"cited_title":"Zhang, M","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional diffusion recipe for single-arm wrist-camera view augmentation that D-CODA extends to two arms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the action-chunking transformer policy that serves as the downstream evaluation backbone."},{"cited_title":"James, Z","cited_arxiv_id":null,"evidence_quote":"Provides the simulation environment and waypoint-based demonstration generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the view-invariant baseline using fine-tuned novel view synthesis, which D-CODA is compared against."}],"review_version":1}