{"id":"a6c20fc5-8af5-40d2-a752-b0fdcc89b137","arxiv_id":"2605.29407","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A FiLM-conditioned ACT policy with multi-modal phase prediction raises dual-arm T-shirt hanging success from 56% to 87% via autonomous failure recovery.","lead":"This paper develops a phase-conditioned imitation learning system using FiLM modulation and a multi-modal phase predictor to let robots handle deformable objects like T-shirts while automatically recovering from failures that vision alone cannot detect. A smart generalist might read it to see how adding force sensing and task-phase awareness can make learned robot policies more reliable for everyday manipulation tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"The 31-point success-rate gain rests on unquantified real-time accuracy of the multi-modal phase predictor for detecting vision-invisible contact failures.","rationale":"The reader’s weakest_assumption directly identifies the same unverified component that carries the closed-loop claim. Because the provided abstract contains no quantitative predictor metrics, the concern remains load-bearing even after the full manuscript is consulted; the FiLM ablations do not substitute for predictor validation. No other internal inconsistency appears in the abstract-level argument.","tokens_in":1719,"tokens_out":373,"duration_ms":8790,"concrete_test":"From the results or supplementary material, extract the phase-prediction accuracy (or F1) on a test set of at least 50 trajectories that include both successful and failure cases; also report the fraction of recovery triggers that were caused by force/pose cues rather than vision. If accuracy < 85% or if >30% of recoveries are vision-only, recompute the success-rate delta after disabling the non-visual channels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result attributes the jump from 56% to 87% hanging success to autonomous recovery triggered by the phase predictor fusing vision, force, and pose. The abstract states that this predictor “detects contact failures that are invisible to vision alone,” yet supplies no phase-classification accuracy, confusion matrix, or failure-detection precision/recall on held-out sequences. If the predictor’s real-time error rate is high, the observed improvement could be driven instead by the hybrid impedance controller or by the FiLM-conditioned policy itself rather than by closed-loop recovery. The t-SNE analysis and FiLM ablations address feature separation but do not directly validate the predictor’s failure-detection reliability under the exact conditions of the dual-arm T-shirt task.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a closed-loop hierarchical framework for deformable object manipulation that augments Action Chunking with Transformers (ACT) via FiLM conditioning on task phase, a multi-modal phase predictor fusing vision/force/pose to detect failures invisible to vision, a hybrid impedance controller, and haptic teleoperation. It claims this mitigates Markovian state aliasing in standard imitation learning and enables autonomous recovery, with the central empirical result being an increase in dual-arm T-shirt hanging success from 56% to 87%.","tokens_in":1873,"tokens_out":385,"duration_ms":19338,"significance":"If the success-rate gain can be rigorously attributed to the phase predictor's failure detection rather than the FiLM policy or impedance controller, the work would provide a concrete mechanism for closed-loop recovery in contact-rich deformable manipulation without requiring explicit failure-mode engineering.","major_comments":[{"comment":"Abstract: The 56% to 87% hanging success improvement is explicitly attributed to autonomous recovery triggered by the multi-modal phase predictor detecting contact failures invisible to vision, yet the manuscript supplies no phase-classification accuracy, confusion matrix, precision/recall, or real-time failure-detection metrics on held-out sequences or during the dual-arm task. Without these, it is impossible to confirm that the predictor (rather than FiLM conditioning or the hybrid controller) drives the reported gain.","section":"Abstract"},{"comment":"Ablation studies and t-SNE analysis: These demonstrate that FiLM induces better-separated phase-specific features than unconditioned or token-level baselines, but they do not isolate or quantify the phase predictor's contribution to failure recovery under the exact conditions of the T-shirt hanging task.","section":"Ablation studies / t-SNE analysis"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive feedback. We respond to each major comment below, indicating where the manuscript will be revised for clarification while maintaining the integrity of the reported results.","responses":[{"response":"We acknowledge that the manuscript does not include direct quantitative metrics (accuracy, confusion matrix, precision/recall) for the phase predictor on held-out data. The 56% baseline reflects the standard ACT policy without phase conditioning or the multi-modal predictor, while the 87% result is achieved only with the full closed-loop system that uses the predictor to detect vision-invisible contact failures and trigger recovery. The multi-modal fusion (vision/force/pose) is specifically motivated by the need to resolve state aliasing in contact-rich phases. In revision we will rephrase the abstract to attribute the gain to the integrated closed-loop framework rather than the predictor in isolation, and we will add a limitations paragraph noting the absence of isolated predictor metrics.","revision_made":"partial","referee_comment":"[Abstract] Abstract: The 56% to 87% hanging success improvement is explicitly attributed to autonomous recovery triggered by the multi-modal phase predictor detecting contact failures invisible to vision, yet the manuscript supplies no phase-classification accuracy, confusion matrix, precision/recall, or real-time failure-detection metrics on held-out sequences or during the dual-arm task. Without these, it is impossible to confirm that the predictor (rather than FiLM conditioning or the hybrid controller) drives the reported gain."},{"response":"The ablation studies and t-SNE visualizations evaluate the benefit of FiLM conditioning on policy feature separation and phase-specific behavior. The predictor's contribution to failure recovery is evidenced indirectly through the end-to-end task results: the 31% absolute improvement occurs only when the multi-modal predictor is active and able to trigger recovery trajectories. No separate ablation that disables only the predictor (while retaining FiLM and the impedance controller) was performed. In the revised manuscript we will add an explicit discussion clarifying that the ablations target the policy encoder while the predictor's role is demonstrated via the recovery mechanism in the full dual-arm task.","revision_made":"partial","referee_comment":"[Ablation studies / t-SNE analysis] Ablation studies and t-SNE analysis: These demonstrate that FiLM induces better-separated phase-specific features than unconditioned or token-level baselines, but they do not isolate or quantify the phase predictor's contribution to failure recovery under the exact conditions of the T-shirt hanging task."}],"tokens_in":1381,"tokens_out":567,"duration_ms":27417,"standing_objections":["Phase-classification accuracy, confusion matrix, precision/recall, or real-time failure-detection metrics for the multi-modal phase predictor on held-out sequences"]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a hierarchical setup that conditions an ACT policy on task phase via FiLM and adds a vision-force-pose predictor to spot failures and trigger recovery. This directly tackles the Markovian limitation in standard imitation learning for deformable objects where similar visuals can demand different actions.\n\nWhat works is the ablation evidence that FiLM modulation beats both the plain ACT baseline and token-level conditioning, plus the t-SNE plots showing cleaner phase separation in feature space. The physical experiments on hanging and removing a T-shirt with two arms are the right kind of test, and the jump from 56% to 87% success is large enough to notice.\n\nThe soft spot is exactly the one in the stress-test note. The abstract credits the autonomous recovery for the improvement and claims the predictor catches contact failures invisible to vision, yet it gives no phase-classification accuracy, precision-recall, or confusion matrix on held-out data. Without those numbers it is impossible to know whether the closed-loop part is actually doing the work or whether the hybrid impedance controller and the conditioned policy alone would have produced most of the gain. The t-SNE and ablation results address feature quality but not real-time failure detection reliability under the exact task conditions.\n\nThis is for groups working on imitation learning for manipulation, especially anyone dealing with deformable objects or contact-rich tasks. The idea is practical and the hardware validation is there, so it deserves a serious referee even though the predictor validation needs to be strengthened before the central claim can be taken at face value.","headline":"The FiLM-conditioned ACT plus multi-modal phase predictor gives a clear 31-point lift on dual-arm T-shirt hanging, but the gain is hard to attribute without any accuracy numbers on the predictor itself.","tokens_in":2348,"tokens_out":393,"would_cite":false,"duration_ms":13730,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Phase-conditioned imitation learning with multi-modal feedback enables autonomous recovery from failures in deformable object tasks, raising hanging success from 56% to 87%.","keywords":["imitation learning","deformable object manipulation","phase conditioning","failure recovery","dual-arm robotics","force feedback","ACT policy"],"falsifier":"An experiment that disables or degrades the multi-modal phase predictor while keeping the rest of the system fixed and measures whether the hanging success rate falls back near 56%.","tokens_in":2630,"feed_emoji":"🤖","tokens_out":590,"duration_ms":21549,"temperature":0.7,"pith_summary":"The paper introduces a closed-loop hierarchical architecture that conditions an imitation learning policy on task phase to handle state aliasing in deformable object manipulation. A FiLM-conditioned ACT encoder produces phase-specific behaviors from a single unified policy while a multi-modal predictor fuses visual, force, and pose signals to estimate phase and detect contact failures invisible to vision. This setup triggers recovery trajectories autonomously and uses hybrid impedance control for compliant execution. Validation on dual-arm T-shirt hanging and removal demonstrates the improvement over standard Markovian policies. Ablation studies confirm that FiLM modulation creates separated feature representations for different phases.","feed_headline":"Phase conditioning lifts T-shirt hanging success from 56% to 87%","feed_subtitle":"Multi-modal predictor detects vision-invisible contact failures and triggers autonomous recovery in dual-arm tasks.","key_machinery":"FiLM-conditioned ACT encoder modulated by a multi-modal phase predictor that fuses visual, force, and pose feedback to estimate current task phase and trigger recovery.","core_discovery":"The authors claim that conditioning an Action Chunking with Transformers policy on estimated task phase via FiLM, combined with real-time multi-modal phase prediction from vision, force, and pose, produces a single policy that executes phase-specific actions and recovers autonomously from execution failures that vision alone cannot detect.","pith_inferences":["The approach could reduce the need for separate policies per phase in other multi-stage robotic tasks.","Force and pose feedback may help in environments where visual occlusion is common.","Similar phase conditioning might address state aliasing in other imitation learning methods beyond ACT."],"forward_implications":["A single policy can produce phase-specific behaviors while sharing action dynamics across phases.","Contact failures invisible to vision can be detected and trigger autonomous recovery trajectories.","FiLM modulation yields well-separated phase-specific feature representations confirmed by t-SNE.","Hybrid impedance control supports compliant execution during both normal and recovery actions."],"fun_headline_variants":["Phase conditioned ACT enables autonomous failure recovery","FiLM conditioned encoder for phase specific ACT behaviors","Multi modal phase predictor triggers autonomous recovery trajectories","T-shirt hanging success reaches 87 percent with phase conditioning","FiLM modulation induces phase specific feature representations in ACT"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The multi-modal phase predictor can accurately estimate the current task phase in real time and reliably detect contact failures that are invisible to vision alone.","fun_headline_variants_meta":{"raw":{"variants":["Phase conditioned ACT enables autonomous failure recovery","FiLM conditioned encoder for phase specific ACT behaviors","Multi modal phase predictor triggers autonomous recovery trajectories","T-shirt hanging success reaches 87 percent with phase conditioning","FiLM modulation induces phase specific feature representations in ACT"]},"model":"grok-4.3","cost_usd":0.008571,"raw_usage":{"total_tokens":3865,"prompt_tokens":658,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":85712000,"prompt_tokens_details":{"text_tokens":658,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3146,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":658,"tokens_out":61,"duration_ms":24207,"temperature":1.0,"reasoning_tokens":3146,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T07:11:44.936036+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment that disables or degrades the multi-modal phase predictor while keeping the rest of the system fixed and measures whether the hanging success rate falls back near 56%.","supporting_citations":[],"review_version":1}