{"id":"d81dd7cf-05d6-470f-93f8-f340ad9212c8","arxiv_id":"2608.13438","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ContactGuard uses an action-conditioned latent world model to predict a post-contact visual embedding from a planned action chunk and abort before the gripper closes, improving failure prediction over current-latent and direct baselines in real-robot grasping.","lead":"A robot monitoring system that predicts whether an upcoming grasp will fail before the gripper closes, using a latent world model that imagines the near future from the planned actions. It can veto a poor grasp half a second before contact, potentially reducing damage and improving reliability in contact-rich manipulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rollout fidelity to the true post-contact latent is never measured, so the claimed 'predicted consequence' mechanism is not established; the probe may succeed despite, not because of, faithful world-model forecasting.","rationale":"The reader's weakest assumption is that the mean-pooled multi-view latent preserves outcome-relevant contact information after being rolled forward. My concern sharpens this into a directly testable precondition: the paper never verifies that the autoregressively predicted latent actually matches the real post-contact latent. High probe AUC on the predicted latent only shows that \\\\hat{z}_{t+K} is a useful feature; it does not show that the world model is faithfully forecasting the true future latent. The current-latent and corrupted-action ablations establish action sensitivity but not predictive fidelity, so the mechanism behind the improvement remains underspecified. This is a correctness risk rather than a stylistic objection, and it can be settled by a concrete offline measurement. The existing conditional verdict already acknowledges that reproducibility and verification are incomplete; adding this fidelity check as a condition does not change the overall verdict, so I recommend UNCHANGED rather than a move to accept or reject.","tokens_in":12714,"tokens_out":10598,"duration_ms":122351,"concrete_test":"On the held-out offline pools, for every labelled clip compute the frozen encoder's real latent at the readout time z_{t+K} and the model's autoregressive prediction \\\\hat{z}_{t+K}. Report mean squared error and cosine similarity between the two, and retrain the same logistic probe on the real future latents using the same CV protocol. If the predicted-to-real similarity is low, or if the real-latent probe's AUC is substantially above ContactGuard's, the forecast is not faithful and the stated mechanism is unsupported; if they are close, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ContactGuard forecasts the post-contact latent and that this imagined future is the source of the failure signal. The paper never reports how close the autoregressively rolled-out latent \\\\hat{z}_{t+K} is to the real latent z_{t+K} on held-out clips, and it never compares the probe's AUC when fed the real future latent instead of the predicted one. Because the world model is trained only on teacher-forced one-step next-latent MSE with SIGReg (Eq. 4), it is not guaranteed, and not shown, that multi-step autoregressive rollouts stay near the true post-contact representation. The current-latent and corrupted-action ablations show the predicted latent carries action-dependent signal, but a corrupted rollout can still produce a useful-yet-unfaithful representation (e.g., an action-embedding shortcut), which would support a narrower claim—an action-conditioned nonlinear probe—rather than the stated latent world-model monitor. Section 4.3 discusses what information the monitor uses only through AUC changes, not through latent-space fidelity; no section verifies that the predicted future latent corresponds to the actual post-contact scene.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ContactGuard, a pre-contact execution monitor for chunked visuomotor policies. It trains a JEPA-style latent world model on unlabelled robot trajectories to predict multi-view visual latents under planned actions, then trains a linear logistic probe on the predicted post-contact latent to score grasp failure. At deployment, the monitor triggers on an imminent gripper-closure event, rolls the frozen predictor forward under the policy's action chunk, and aborts if the predicted failure probability exceeds a per-task threshold. The paper reports four real-world grasp tasks (cup, box, pencil, towel) and compares against a single-view world model, direct classification from the current latent plus actions, a matched current-latent ablation, and external failure detectors. The main empirical claims are that predicted future latents yield higher ROC AUC than current-latent and direct baselines, that action corruption destroys the signal, and that the system runs fast enough for a 0.5 s pre-contact window. The paper also includes a live real-robot evaluation with abort-based counterfactual outcome recording.","tokens_in":12932,"tokens_out":7805,"duration_ms":76740,"significance":"If the results hold, ContactGuard is a practical contribution: it adds a policy-decoupled predictive verifier to chunked visuomotor policies without retraining or candidate search. The ablation design in Table 2 is a genuine strength: the matched current-latent baseline shares the encoder, probe class, data pools, and cross-validation protocol, and the reported five-seed cross-validation makes the offline comparison credible. The action-swap test in Section 4.3 is a particularly clean falsifiable intervention. However, the paper's mechanistic claim—that the probe reads a faithful forecast of the post-contact latent—is not yet supported, and one baseline is not fully matched on input context. If the missing controls are added and support the current results, this would be a solid systems contribution to runtime monitoring.","major_comments":[{"comment":"The central mechanistic claim is that the failure signal comes from a faithful prediction of the post-contact latent, but no fidelity check is reported. The paper never compares the autoregressively rolled-out \\hat{z}_{t+K} with the true latent z_{t+K} on held-out clips, and it never reports the probe's AUC when fed the true future latent instead of the predicted one. Because the world model is trained only with teacher-forced one-step MSE plus SIGReg (Eq. 4), multi-step rollouts are not guaranteed to stay near the true latent. The current-latent and corrupted-action ablations show that the signal is action-dependent, but a corrupted or unfaithful rollout could still produce a useful action-conditional representation. Add a latent-fidelity measurement (e.g., MSE or cosine similarity between predicted and true future latents at K=1, 15, 30) and a probe-on-real-future-latent control; without these, the abstract's claim that the monitor 'predicts its short-horizon consequence in latent visual space' is not established.","section":"§3.3, §4.3, Eqs. (3)–(4)"},{"comment":"The 'Current latent' ablation is not fully matched on input features. ContactGuard's probe reads \\hat{z}_{t+K}, which is a function of the C=3 latent context z_{t-C+1:t} and K future actions, whereas the Current latent baseline reads only the single anchor latent z_t. The observed AUC gain could therefore be due in part to the additional temporal context rather than to the imagined future. Add a baseline that consumes the full context without rollout—for example, the mean or concatenation of z_{t-C+1:t} fed to the same linear probe—to isolate the contribution of the action-conditioned latent prediction.","section":"§4.2, Table 2; §4.3, Table 3"},{"comment":"The live closed-loop results in Table 1 are reported as point estimates without any uncertainty quantification. With N=50 per task, the difference between Ours and LeWM on Box (AUC 0.946 vs 0.933, FAR 0.107 vs 0.107) and on Towel (balanced accuracy 0.820 vs 0.700) could be within sampling noise; for example, a single confusion-matrix entry change on Box alters recall by about 0.04. Report bootstrap confidence intervals or repeated-seed statistics for the live metrics, or at minimum state the number of independent rollouts per cell and how thresholds were fixed. This is needed to support the paper's claim that the monitor 'transfers to live robot'.","section":"§4.2, Table 1"}],"minor_comments":[{"comment":"Both Table 2 and Table 3 report a 'Current latent' row with different values (e.g., Cup 0.840 vs 0.758) because they use different offline splits. The text says this in passing in §4.3, but the discrepancy should be flagged explicitly in the captions to avoid confusion.","section":"Table 2, Table 3"},{"comment":"Figure 1 uses 'h steps' for the rollout horizon while the text and equations use K; unify the notation.","section":"Figure 1"},{"comment":"The abstract has several formatting artifacts (e.g., missing space after the ContactGuard emphasis, 'setups' typo); please run a clean proofread and LaTeX pass.","section":"Abstract"},{"comment":"The text states the anchor is 0.5 s before closure and the rollout reaches 0.5 s after closure at 30 Hz, but the phrase could be misread because k_pre=15 frames at 30 Hz is 0.5 s and K=30 steps reaches T_g + 15 frames. Clarify the exact frame indices.","section":"§4.1"},{"comment":"The class-weight notation is slightly confusing: s_c is defined for class c, but the loss uses s_{y_i}; consider defining the weight directly as a function of the sample's class to improve readability.","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for the venue and the empirical contribution is likely salvageable. My main concern is that the mechanistic interpretation as a faithful world model is not yet probed, and the 'Current latent' baseline conflates temporal context with imagined future. I recommend major revision rather than rejection because the core comparison is well designed and the requested controls are well within the scope of a revision. No citation or novelty concerns beyond those already discussed in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Over coffee: this is a genuinely useful paper. ContactGuard takes a JEPA-style latent world model, trains it on unlabelled trajectories, and uses a linear probe on the predicted post-contact latent to abort before gripper closure. The real-robot evidence is the strongest part: four grasping tasks, fifty live rollouts per task, and the monitor beats both a direct classifier and a capacity-matched single-view baseline. The 'Current latent' ablation is well done—it uses the same encoder, probe family, and CV protocol, so the AUC gap is fairly attributed to the action-conditioned rollout. The action-swap counterfactual, where swapping only the planned action changes the failure score, is a nice addition that shows the monitor is not just reading static visual risk.\n\nWhat's new is the policy-decoupled interface: the monitor treats the policy as a black box, consumes the emitted action chunk, and vetoes it before contact. That's a practical niche that SIRIUS and similar joint-training monitors don't fill. The writing is honest—it explicitly disclaims novelty on imagined-future monitoring and positions the contribution as the deployment interface.\n\nSoft spots: the stress-test note is right that the paper never verifies rollout fidelity. There is no measurement of how close the autoregressively predicted latent is to the true post-contact latent, and no probe-on-real-future-latent comparison. Without that, the stronger mechanism claim—that the world model forecasts the consequence—isn't fully established. The action-corruption and swap results mitigate the harshest version of the critique, because the signal is clearly action-dependent rather than a pure visual shortcut, but they don't pin down that the latent is a faithful model of the scene. This is a missing experiment rather than a fatal flaw; the practical claim survives.\n\nOther issues are minor: no code or data release, and the live results in Table 1 lack error bars. The per-task thresholds are fitted on validation, which is standard but worth noting. The paper acknowledges its own limitations (no recovery after abort, short-horizon events) without overselling.\n\nWho gets value: robotics people working on runtime safety for learned policies, and anyone building monitors for action-chunked imitation learning. It is not a new scientific framework, but it is a well-validated system with clean comparisons.\n\nRecommendation: it deserves a serious referee. I would send it out, but I would ask the authors to either measure latent fidelity (predict vs. real future latent, or probe on the real latent) or explicitly narrow the claim to 'action-conditioned probe on a learned latent rollout' rather than 'world model faithfully forecasts the post-contact scene.' Release of code and data would also help.","headline":"Solid, real-robot pre-contact failure monitor with clean ablations; the one missing check is whether the latent rollout actually tracks the post-contact scene.","tokens_in":13458,"tokens_out":2960,"would_cite":true,"duration_ms":30601,"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":"ContactGuard predicts grasp failures before contact by rolling the planned action chunk through a frozen latent world model and vetoing the grasp when the predicted post-contact latent flags failure.","keywords":["pre-contact execution monitoring","latent world models","action-conditioned prediction","grasp failure prediction","visuomotor policies","JEPA","multi-view fusion"],"falsifier":"Collect a labeled set in which the same planned action chunk is executed from the same observation but succeeds and fails due to unobserved object state; if ContactGuard's predicted post-contact latent cannot separate these cases (ROC AUC near 0.5), the imagined-future signal is not carrying the outcome information the paper claims.","tokens_in":12509,"feed_emoji":"🤖","tokens_out":7916,"duration_ms":77749,"temperature":0.7,"pith_summary":"ContactGuard is a pre-contact execution monitor for chunked visuomotor policies. It takes the action chunk the policy is about to execute, rolls it forward through a frozen latent world model trained on unlabelled robot trajectories, and uses a small linear probe on the predicted post-contact latent to decide whether the grasp will fail. The paper's central claim is that this imagined future latent carries failure information not available from the current observation or the planned action alone, and that the monitor can veto a bad grasp before the gripper closes on a live robot without changing the policy. Across four real-world tasks, the predicted future latent improves ROC AUC over a matched current-latent baseline, and action-corruption experiments show the signal is tied to the specific proposed action. The paper is explicit that this is a veto, not a recovery: aborting stops the failure but does not complete the task, and a separate recovery module is left as future work.","feed_headline":"World-model monitor aborts bad grasps before contact","feed_subtitle":"It rolls the action into latent space and vetoes the chunk if the probe flags failure, across four real-robot tasks.","key_machinery":"The central object is the action-conditioned latent world model: a shared ViT-Tiny encoder maps each camera view to a compact embedding, the per-view embeddings are mean-pooled across cameras into one latent, and a causal Transformer predictor (conditioned on actions through AdaLN-zero modulation and regularized with SIGReg) is trained with next-latent regression to roll the latent forward under planned actions. This follows the Joint-Embedding Predictive Architecture (JEPA) principle, predicting the embedding of a future signal rather than reconstructing pixels. At deployment the predictor is unrolled autoregressively for 30 steps from an anchor 15 frames before gripper closure, and a frozen ℓ2-regularized logistic regression probe scores the final predicted latent. This machinery converts a concrete proposed action chunk into a predicted visual-consequence latent that a low-capacity linear readout can classify, without pixel-level video prediction.","core_discovery":"On its own terms, ContactGuard establishes that an action-conditioned latent world model, trained without any success labels, produces a representation of a planned grasp's near future that a lightweight logistic probe can read to separate successful from failed grasps. Across four real tasks (cup and box pick-and-place, pencil-and-notebook, towel-fold), the predicted future latent achieves higher ROC AUC than the matched current-latent baseline and than external failure detectors, and live robot rollouts show the monitor aborting before closure with lower false-abort rates than direct prediction from the anchor latent and action. The paper also shows the signal is action-specific: holding the observation fixed and swapping in an action chunk from a failed attempt increases the predicted failure probability, while corrupting the action chunk collapses performance to near chance.","pith_inferences":["The trigger in this paper is a gripper open-to-close transition, but the underlying rollout-and-probe machinery is trigger-agnostic; an equivalent monitor could be built for other imminent contact events such as insertion, pressing, or pushing by only changing the trigger detector.","Because the world model is trained without outcome labels, its latent is not optimized to retain slip or misalignment details; a future objective that adds an auxiliary task-outcome or contrastive loss might raise the ceiling, or reveal a failure mode on tasks where those details are not visually expressed.","The action-swap result suggests ContactGuard could be used offline as an action-chunk evaluator, screening or ranking candidate chunks before execution rather than only aborting a single chunk at runtime.","The false-abort variation across tasks (for example, towel trading precision for coverage) implies threshold choice is consequential; a user deploying this would want to calibrate the threshold on their own success/failure distribution."],"forward_implications":["ContactGuard can be attached to an existing chunked visuomotor policy as a black-box verifier, so deployment does not require retraining or modifying the policy.","Pre-contact abort decisions happen within the control loop: the full encode–rollout–probe pass at 30 steps takes about 19 ms, leaving slack before gripper closure at 30 Hz.","The monitor's signal comes from the specific pending action, not just a risky-looking scene; corrupting or swapping the action chunk changes the verdict.","Multi-view mean-fused latents improve prediction over any single view, particularly when the contact region is occluded from one camera.","The monitor is a veto, not a recovery; after an abort the task still needs an external recovery module, which the paper leaves as future work."],"supporting_citations":[{"why":"Supplies the JEPA principle of predicting the embedding of a target signal rather than reconstructing it in pixel space.","marker":"[1]"},{"why":"Provides the LeWorldModel latent predictor and SIGReg regularizer that ContactGuard adapts for multi-view pre-contact monitoring.","marker":"[11]"},{"why":"The chunked visuomotor policy whose action chunks ContactGuard monitors in the experiments.","marker":"[30]"},{"why":"A success-only runtime failure detector used as a comparison baseline for pre-contact prediction.","marker":"[18]"},{"why":"Provides a novelty-based success-only reference signal for failure prediction.","marker":"[19]"},{"why":"A supervised multi-task failure detector used as an external supervised baseline.","marker":"[20]"},{"why":"Closest prior work in imagined-state runtime monitoring; the paper contrasts its black-box proposer interface with this joint policy/world-model training approach.","marker":"[23]"},{"why":"Prior work on early grasp-failure prediction with sequence models; ContactGuard extends this to prediction before any contact by rolling the action chunk.","marker":"[25]"}],"fun_headline_variants":["Latent world model predicts grasp failure before contact","Pre-contact monitor rolls actions into latent space to veto failures","Action-conditioned world model aborts bad grip plans pre-contact","Wrist-cam latent world model halts failed grasps pre-contact","Latent rollouts foresee grasp failure and abort before contact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system relies on the compressed visual snapshot averaged across cameras, after being rolled forward by the world model, preserving enough information about contact (slip, displacement, misalignment) to tell success from failure half a second after the gripper closes, even though the world model is never trained on outcome labels.","fun_headline_variants_meta":{"raw":{"variants":["Latent world model predicts grasp failure before contact","Pre-contact monitor rolls actions into latent space to veto failures","Action-conditioned world model aborts bad grip plans pre-contact","Wrist-cam latent world model halts failed grasps pre-contact","Latent rollouts foresee grasp failure and abort before contact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000415,"raw_usage":{"total_tokens":2124,"prompt_tokens":907,"completion_tokens":1217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1134}},"tokens_in":523,"tokens_out":1217,"duration_ms":10226,"temperature":1.0,"reasoning_tokens":1134,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:59:33.201622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a labeled set in which the same planned action chunk is executed from the same observation but succeeds and fails due to unobserved object state; if ContactGuard's predicted post-contact latent cannot separate these cases (ROC AUC near 0.5), the imagined-future signal is not carrying the outcome information the paper claims.","supporting_citations":[{"cited_title":"Assran, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the JEPA principle of predicting the embedding of a target signal rather than reconstructing it in pixel space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A success-only runtime failure detector used as a comparison baseline for pre-contact prediction."},{"cited_title":"Burda, H","cited_arxiv_id":null,"evidence_quote":"Provides a novelty-based success-only reference signal for failure prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A supervised multi-task failure detector used as an external supervised baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Closest prior work in imagined-state runtime monitoring; the paper contrasts its black-box proposer interface with this joint policy/world-model training approach."},{"cited_title":"Damak, M","cited_arxiv_id":null,"evidence_quote":"Prior work on early grasp-failure prediction with sequence models; ContactGuard extends this to prediction before any contact by rolling the action chunk."}],"review_version":1}