{"id":"f00a61c4-d083-4bb3-b564-707c2e09fe3b","arxiv_id":"2505.06628","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"ACORN adds contrastive learning with action-space perturbations to ACT imitation learning and reports improved safety metrics on one bimanual task, but the headline metric is undefined and results are inconsistent across noise levels.","lead":"This paper introduces ACORN, a plug-in training routine that adds contrastive learning to the ACT robotic manipulation policy, using noisy action samples as negatives to push behavior away from unsafe trajectories. The authors report up to 23% higher failure-conditional reward under moderate action noise, but the reward is never defined and the improvement disappears under light noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ACR-F and ACR depend on a reward r_t that is never defined for an imitation-learning benchmark, making the 23.1% headline improvement uninterpretable.","rationale":"I agree with the reader's weakest assumption and the REJECT verdict. The single most load-bearing point is that Eq. (3) uses r_t without definition or source. The ACT/ACORN training objective in Eq. (1) is supervised regression plus KL regularization, and the LeRobot AlohaTransferCube task is an offline imitation-learning benchmark that does not provide reward labels. Therefore ACR and ACR-F, including the 23.1% Normal-noise ACR-F gain in Sec. V-B, are not computable from the described protocol. An arbitrary reward function could produce almost any ordering, so the central claim is currently unsupported rather than merely weakly supported. The secondary concern that the training negative perturbations (Eq. 7) use the same Gaussian scaling family as the test perturbations (Eq. 6) would further weaken the robustness interpretation, but the undefined reward is the decisive issue. No machine-checked proofs or released evaluation code are offered to pin down r_t. The proposed check would settle the concern by recovering the exact reward and recomputing the headline numbers.","tokens_in":8078,"tokens_out":2862,"duration_ms":31846,"concrete_test":"Ask the authors for the evaluation code or exact reward specification used to compute ACR and ACR-F; then re-run the Normal-noise evaluation of ACT and ACT-ACORN on AlohaTransferCube with that explicit r_t and the same 500 episodes and failure set. If no reward signal exists in LeRobot's imitation pipeline, ACR-F cannot be computed and Table I's reward rows are vacuous; if a reward is supplied, verify that the 86.116 vs 69.952 comparison (and the +11.8% Heavy ACR-F gain) reproduces. Also check whether reward values are bounded or normalized, since Eq. (3) sums raw rewards over variable episode lengths.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is the Normal-noise ACR-F gain (86.12 vs 69.95, Sec. V-B). ACR-F is defined in Eq. (3) as 1/|F| times the sum over failed episodes of sum_t r_t^(i), but no reward function r_t is defined anywhere in the paper. The ACT baseline and the ACORN extension are trained with the supervised ACT objective in Eq. (1), which is a regression plus KL objective, and the LeRobot AlohaTransferCube task is an imitation-learning benchmark with no reward signal. Consequently, the headline metric cannot be computed from the stated setup, and the reported 23.1% improvement is not a well-defined comparison. The same gap affects ACR, which is also an average cumulative reward, and therefore affects the other reward-based rows of Table I. This is load-bearing: without a specified r_t, the direction and magnitude of the ACR-F comparison are unconstrained by the paper. An arbitrary reward could reverse or erase the claimed improvement, and a reader cannot verify that ACORN improves safety rather than merely shifting an unstated scoring rule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACORN, a plug-and-play extension of the Action Chunking with Transformers (ACT) imitation learning policy. ACORN adds contrastive negative sampling via correlated scaling and independent micro-perturbations, a curriculum-scheduled contrastive loss, and a Huber regression term. The paper also introduces four safety metrics (ACR-F, AM-J, AM-E, TDL) and reports experiments on the LeRobot AlohaTransferCube task under Light, Normal, and Heavy multiplicative Gaussian action noise, claiming up to 23% improvements in ACR-F without sacrificing success rate.","tokens_in":8337,"tokens_out":3895,"duration_ms":41908,"significance":"If the claims held, ACORN would be a practically valuable robustness wrapper for ACT-style policies, since it is architecture-compatible and requires no additional data collection. The paper also contributes a plausible evaluation template that perturbs actions at three noise levels, and it reports ablations for the curriculum coefficient k and the contrastive margin alpha. However, the headline claim depends on an undefined reward function, the evaluation protocol overlaps with the training-time noise family, and no variance or multi-seed statistics are reported. The central safety claim is therefore not currently supported, and the contribution remains an idea plus unverified numbers rather than a validated method.","major_comments":[{"comment":"ACR-F is defined as the average cumulative reward over failed episodes, but no reward function r_t^(i) is defined anywhere in the manuscript. The ACT objective in Eq. (1) is supervised imitation (L1 reconstruction plus KL regularization), and the LeRobot AlohaTransferCube task is an imitation-learning benchmark with no reward signal. Consequently, the Normal-noise ACR-F comparison in Table I (86.116 vs 69.952) and the abstract's 23% improvement claim cannot be computed or independently verified. This is load-bearing because ACR and ACR-F are used as primary evidence of safety improvement; the paper must either define the reward used in evaluation or remove reward-based metrics.","section":"Section IV-A, Eq. (3); Table I"},{"comment":"No variance, confidence intervals, or multiple seeds are reported. The text states that 500 episodes per noise condition were evaluated, but each entry in Table I is a single point estimate. Given the stochasticity of the noise injection in Eq. (6) and the high variance of manipulation success rates (SR ranges from 22.8% to 80.8% across conditions), small differences such as SR 22.80 vs 23.00 or AM-E 0.116 vs 0.115 are indistinguishable from sampling noise. To support the claimed improvements, the authors should report mean and standard deviation over at least 3-5 independent training or evaluation seeds.","section":"Section V-A, Table I"},{"comment":"The evaluation noise family matches the training-time negative sample distribution. Eq. (6) generates perturbed actions as (1+beta) a_gt with beta = eta*sigma and eta ~ N(0,1), i.e., multiplicative Gaussian scaling. Eq. (7) generates negative samples as (1 + eta*sigma) * a_gt + epsilon, which is the same multiplicative Gaussian family plus an independent micro-perturbation. ACORN is therefore evaluated on the perturbation distribution it was explicitly trained against, while the ACT baseline was not exposed to this distribution. This confounds the robustness comparison and does not demonstrate generalization to realistic disturbances. A concrete fix is to evaluate on held-out noise families (additive Gaussian, uniform, sensor noise) or to train the baseline with the same negative-sample augmentation.","section":"Section IV-B, Eq. (6); Section IV-C, Eq. (7)"},{"comment":"AM-J and AM-E measure deviation from the nearest training demonstrations, and ACORN's contrastive loss explicitly minimizes distance to expert positive samples while ACT only regresses with L1/Huber loss. Lower AM-J/AM-E under noise is therefore partly expected by construction and does not, by itself, establish safer behavior. The paper should validate these kinematic metrics against independent safety indicators, such as workspace-boundary or obstacle violations, and should report TDL quantitatively rather than only as a visualization concept, since Table I contains no TDL values.","section":"Section IV-A, Eqs. (4)-(5); Section V-B"}],"minor_comments":[{"comment":"The text contains a duplicated phrase: 'with with 1) Huber Loss and 2) Contrastive Margin Loss'.","section":"Section IV, Stage II"},{"comment":"The sentence 'ACORN is purposed as a plug-and-play optimization to exhibit catastrophic failures under distributional shifts' appears to state the opposite of the intended meaning; it should read 'prevent' rather than 'exhibit'.","section":"Section IV"},{"comment":"The notation lambda_KL L_KLD is introduced without definition; Eq. (1) defines lambda and D_KL. Please align the notation and define every symbol in Eq. (8).","section":"Eq. (8)"},{"comment":"TDL is introduced as one of the four safety metrics but is never reported numerically in the experiments; please specify how confidence regions and violations are computed and provide quantitative results.","section":"Section IV-A, TDL"},{"comment":"The claims about 'more concentrated region of trajectory density' and 'substantial reduction in trajectories violating safety thresholds' are not supported by any quantitative trajectory-density or violation counts; please provide numerical measures.","section":"Section V-B, Figs. 3-4"}],"recommendation":"reject","confidential_remarks":"The manuscript is not reproducible in its current form because the reward function behind the headline metric is unspecified and no code or data release is indicated. The evaluation also risks being circular because the test-time noise family coincides with the training-time negative-sample distribution. If the authors define a principled reward (or replace ACR/ACR-F with well-defined metrics), report multi-seed statistics, and evaluate on held-out perturbation families, a substantially revised resubmission could merit reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nACORN is a contrastive-learning add-on to ACT that tries to make imitation policies robust to action noise. The core idea is reasonable: generate noisy negative action trajectories, then push the policy away from them with a contrastive loss while keeping it close to expert actions. The paper also adds a curriculum schedule for the contrastive weight and a Huber loss. That is a sensible combination, and the ablation studies for k and α are actually informative.\n\nThe problem is that the headline result—a 23.1% improvement in ACR-F under Normal noise—is not something you can check. ACR-F, defined in Eq (3), is an average cumulative reward over failed episodes. But no reward function r_t is defined anywhere. The task is supervised imitation learning on AlohaTransferCube; there is no reward signal in that setup. So the 86.12 vs 69.95 number is uninterpretable. The same goes for ACR in Table I. This is not a minor omission; it is load-bearing. Without a definition of r_t, the direction of the comparison is unconstrained. A reader cannot verify that ACORN is safer rather than just gaming an unstated scoring rule.\n\nThe other metrics are also weaker than they look. AM-J and AM-E measure deviation from the training demonstrations, which is partly by construction: ACORN is trained to imitate those demonstrations, so lower values are somewhat expected. The evaluation injects multiplicative Gaussian action noise (Eq 6), the same family used to generate the negative samples (Eq 7), which raises a question about overfitting to that noise profile. There are no error bars, no multiple seeds, only one task and one baseline. And the hyperparameters (k=15, α=0.01) were selected via ablation on the Normal noise condition that also appears in the main results—a mild test-set tuning issue.\n\nWhat the paper does well: the method is simple and plausible, the presentation is clear, and the authors are upfront about the architecture and training details. The idea of using contrastive learning to separate expert and noisy action trajectories is worth exploring. But the evidence, as it stands, does not support the claim.\n\nFor peer review: I would send it to referees rather than desk reject, because the core idea has merit and the flaws are fixable. But I would reject in current form and ask for a defined reward, error bars, an additional task, and an evaluation that distinguishes safety from closeness to the training set.\n\nThat's my take. Let's talk over coffee if you want.","headline":"A plausible contrastive robustness add-on for ACT, but the headline ACR-F metric depends on an undefined reward and the evaluation lacks error bars, so the central claim needs better support before it can be trusted.","tokens_in":8807,"tokens_out":2946,"would_cite":false,"duration_ms":28731,"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 new loss makes robot imitation policies safer under noise, lifting failure-conditional reward by 23.1% without hurting success rate.","keywords":["robotic manipulation","imitation learning","Action Chunking with Transformers","contrastive learning","safety metrics","perturbation robustness","curriculum learning","ACORN"],"falsifier":"Replicate the AlohaTransferCube comparison and require the authors to specify $r_t$ and compute ACR-F explicitly; if no reward exists in the imitation-learning setup, the claimed 23.1% gap cannot be reproduced. A simpler check is to rerun the normal-noise evaluation with additional random seeds and see whether the ACR-F difference between ACT and ACT-ACORN persists.","tokens_in":1419,"feed_emoji":"🦾","tokens_out":1531,"duration_ms":75245,"temperature":0.7,"pith_summary":"ACORN claims to be a plug-and-play training objective that extends Action Chunking with Transformers (ACT), a common imitation learning method for fine-grained bimanual manipulation. It adds a contrastive loss that pulls predicted action chunks toward expert demonstrations and pushes them away from synthetically perturbed 'unsafe' action chunks, with a curriculum schedule that grows the contrastive weight as the baseline loss drops. The paper reports that under a normal level of actuator noise, ACT-ACORN matches the baseline success rate (62.2%) while raising the failure-conditional reward ACR-F from 69.95 to 86.12, a 23.1% improvement. If true, ACORN gives a drop-in robustness enhancement that requires no extra demonstrations and no changes to the policy architecture.","feed_headline":"A plug-in loss lifts safety 23% with no success-rate cost","feed_subtitle":"ACORN adds contrastive training so ACT-style policies reject noisy actions without extra data.","key_machinery":"The central object is the ACT-ACORN loss, $$L_{\\text{ACT-ACORN}} = L_{\\text{Huber}} + \\lambda_{\\text{KL}} L_{\\text{KL}}^D + \\lambda_c(L_b) L_{\\text{Contrast}},$$ with $$L_{\\text{Contrast}} = \\max(0, \\|a^+ - \\hat{a}\\|_2 - \\|a^- - \\hat{a}\\|_2 + \\$\\alpha$).$$ Negative action chunks are generated by dual perturbation, $$a^- = (1 + \\eta\\$\\sigma$) \\odot a_{gt} + \\epsilon,$$ combining correlated scaling with independent micro-perturbations. The contrastive weight $\\lambda_c(L_b)$ follows a curriculum schedule driven by the baseline loss, so the policy first learns to imitate and then increasingly learns to reject noisy trajectories. This machinery turns every training batch into paired positive and negative action examples, teaching the policy to hug expert demonstrations while actively avoiding unsafe actions.","core_discovery":"The paper's central claim is that ACORN, an extension of ACT, improves safety and robustness against actuator noise without sacrificing task performance. The mechanism is a composite objective in which the usual ACT losses (an adaptive Huber regression loss and KL regularization) are augmented by a contrastive margin loss that pushes predicted action chunks toward expert demonstrations and away from synthetically corrupted ones. Negative samples are generated by correlated Gaussian scaling plus tiny independent perturbations, and the contrastive weight grows through a curriculum schedule tied to the baseline loss. In the reported AlohaTransferCube evaluation, ACT-ACORN matches the baseline success rates under all tested noise levels (80.80 vs 77.40 for light, 62.20 vs 62.20 for normal, 22.80 vs 23.00 for heavy), while substantially improving ACR-F under normal and heavy noise and reducing AM-E, which the paper interprets as safer, more precise, and more energy-efficient manipulation.","pith_inferences":["Because the contrastive loss operates purely on action chunks, the same negative-sample recipe should transfer to other action-sequence policies beyond ACT, though this is not tested in the paper.","The joint grouping (shoulder, elbow, and forearm roll versus waist and wrist) suggests a testable extension: apply larger contrastive weights to high-priority joints and measure whether safety gains concentrate there.","Until a reward function for ACR-F is specified, AM-J and AM-E are the only directly computable safety metrics in the paper, so reproducing the headline 23.1% improvement requires that missing definition.","The curriculum schedule, with its sharp transition once the baseline loss drops below one, invites sensitivity tests on other tasks, since the exponential coefficient $k$ is tuned on a single environment."],"forward_implications":["Any ACT-style imitation policy could be hardened against action noise by adding dual-perturbation negative samples and a curriculum-scheduled contrastive loss, with no new demonstrations or network changes.","If the reported equal success rate under normal noise holds, robustness improvements need not trade off task completion, at least on the AlohaTransferCube task.","The adaptive Huber loss and the margin parameter provide new tuning knobs whose effects on safety metrics are demonstrated in the ablations.","The four safety metrics, once the reward component is defined, could become a standard way to compare manipulation policies under disturbance.","The biomechanical joint grouping suggests that concentrating contrastive weight on high-priority joints could be a cheap way to target precision improvements, though the paper only visualizes this effect and does not ablate it separately."],"supporting_citations":[{"why":"Supplies the ACT baseline policy and the action-chunking formulation that ACORN extends.","marker":"[1]"},{"why":"Provides the contrastive learning framing and pair-based objective adapted into the action-space margin loss.","marker":"[4]"},{"why":"Gives the curriculum learning concept used to set the contrastive weight $\\lambda_c(L_b)$.","marker":"[28]"},{"why":"Provides the open-source implementation framework, the AlohaTransferCube task environment, and the 50-demonstration dataset used in all experiments.","marker":"[29]"}],"fun_headline_variants":["ACORN: contrastive loss boosts robot safety 23% without hurting success","Contrastive optimization lifts manipulation safety 23% at zero performance cost","Adaptive contrastive trick improves robot safety by up to 23%","ACORN makes robot policies safer under noise, no accuracy loss"],"cache_read_input_tokens":11008,"weakest_assumption_plain":"The load-bearing premise is that the headline safety metric ACR-F is computable, but the paper never defines the reward function $r_t$ that ACR-F sums, and the supervised imitation-learning setup supplies no reward signal, so the 23.1% improvement may rest on an uncomputed quantity.","fun_headline_variants_meta":{"raw":{"variants":["ACORN: contrastive loss boosts robot safety 23% without hurting success","Contrastive optimization lifts manipulation safety 23% at zero performance cost","Adaptive contrastive trick improves robot safety by up to 23%","ACORN makes robot policies safer under noise, no accuracy loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2924,"prompt_tokens":919,"completion_tokens":2005,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":1927}},"tokens_in":535,"tokens_out":2005,"duration_ms":13940,"temperature":1.0,"reasoning_tokens":1927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:37:11.435976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replicate the AlohaTransferCube comparison and require the authors to specify $r_t$ and compute ACR-F explicitly; if no reward exists in the imitation-learning setup, the claimed 23.1% gap cannot be reproduced. A simpler check is to rerun the normal-noise evaluation with additional random seeds and see whether the ACR-F difference between ACT and ACT-ACORN persists.","supporting_citations":[{"cited_title":"A simple frame- work for contrastive learning of visual representations,","cited_arxiv_id":null,"evidence_quote":"Provides the contrastive learning framing and pair-based objective adapted into the action-space margin loss."}],"review_version":1}