{"id":"9c9f64ea-d8ca-4b6b-bbc3-defd45927097","arxiv_id":"2607.06018","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"RoboTALES uses hierarchical LLM subgoals and VLM reward feedback to keep video-model futures task-aligned, then trains robot policies that beat baselines on RoboCasa and LIBERO10 long-horizon tasks.","lead":"RoboTALES trains robot policies from video-model imagined futures that stay on-task via an LLM subgoal planner and a VLM critic. A generalist might care because it shows how language reasoning can stop generative imagination from drifting on long robot tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"VLM critic scores on synthetic futures lack reported calibration against true task progress, so the anti-drift claim and long-horizon policy gains rest on an untested alignment premise.","rationale":"The Reader correctly identified the VLM-on-synthetic accuracy premise as the weakest assumption; after examining the argument structure that remains the single most load-bearing concern. The hierarchical planner is a plausible systems contribution and outperformance on long-horizon RoboCasa/LIBERO10 tasks is the right kind of evidence if the numbers hold, but both the anti-drift claim and causal attribution of gains to “task-aligned simulated futures” require that critic feedback is informative rather than noisy or gamed. The methods description of reward-based feedback does not close this loop with calibration against ground-truth progress or a decisive planner-held-fixed ablation of the critic. I therefore agree with the Reader’s weakest_assumption and would move the verdict from UNVERDICTED to CONDITIONAL: the contribution is plausible and the benchmarks are appropriate, but acceptance should wait on evidence that VLM scores track true progress on the visual domain the generator produces. No internal inconsistency is claimed; this is a missing validation of the key feedback signal that the rest of the pipeline depends on.","tokens_in":1989,"tokens_out":655,"duration_ms":54835,"concrete_test":"On a held-out set of successful and failed trajectories from RoboCasa and LIBERO10, score each trajectory with the same VLM critic used in training (matching the visual format of synthetic futures as closely as possible). Report Spearman ρ / AUROC between VLM reward and binary success or continuous progress. Separately, retrain with the VLM replaced by a progress oracle vs. a shuffled-reward baseline while holding the hierarchical planner fixed; if the VLM does not substantially beat shuffled rewards on final policy success, the alignment premise fails and headline gains need re-attribution.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"For the central claim to hold—that hierarchical LLM subgoals plus VLM reward feedback produce task-aligned simulated futures that yield superior policies—the VLM critic must assign rewards on imagined video that actually track true task progress. Innovation (2) and the abstract’s anti-drift claim depend on this link. The method uses a VLM to score generator rollouts and feed rewards back into training, but the argument does not secure (i) correlation of those scores with ground-truth success or progress on real/sim trajectories in the same visual domain, (ii) robustness when the VLM is biased on synthetic frames, or (iii) an ablation that freezes the planner and replaces the VLM with noisy/constant/shuffled rewards. Without that evidence, gains on RoboCasa and LIBERO10 (especially long-horizon) could come from the hierarchical planner alone, from video-backbone capacity, or from reward-hacking VLM idiosyncrasies rather than genuine task alignment of the futures. That score-fidelity step is the least secure link in the causal chain from “reasoning-guided imagination” to “better policies.”","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"RoboTALES is a single-stage framework that trains robot policies from task-aligned simulated futures produced by a pretrained video generative model. It introduces (1) a hierarchical LLM planner that decomposes tasks into subgoals to condition imagination, and (2) a VLM-based critic that scores imagined rollouts and supplies reward feedback so the generator stays goal-focused and action-conditional. The resulting temporally consistent futures are used to extract policies. The paper reports consistent outperformance over existing methods on RoboCasa and LIBERO10 manipulation benchmarks, with the largest gains on long-horizon tasks, and releases code and models.","tokens_in":2266,"tokens_out":842,"duration_ms":19369,"significance":"If the results hold under scrutiny, the work offers a practical route to make pretrained video generators usable for visuomotor policy learning by coupling hierarchical language reasoning with VLM reward feedback, addressing drift and weak action-conditioning that currently limit imagination-based control. Public code and models support reproducibility. The contribution is primarily empirical and systems-oriented rather than a new theoretical guarantee; its value depends on whether the VLM critic truly aligns synthetic futures with task progress and whether gains are attributable to the claimed mechanisms rather than planner capacity or backbone scale alone.","major_comments":[{"comment":"The central anti-drift and task-alignment claim rests on the VLM critic assigning rewards on synthetic video that track true task progress. The manuscript needs an explicit calibration or correlation analysis (VLM scores vs. ground-truth success/progress on held-out real or sim trajectories in the same visual domain) and a control that freezes the hierarchical planner while replacing VLM rewards with constant, noisy, or shuffled scores. Without this, long-horizon gains on RoboCasa/LIBERO10 cannot be attributed to innovation (2) rather than the LLM planner, video-backbone capacity, or VLM idiosyncrasies.","section":null},{"comment":"Headline outperformance is asserted for RoboCasa and LIBERO10, especially long-horizon tasks, but the evaluation must report full tables with baselines, sample sizes, seeds/error bars, and ablations that isolate (a) hierarchical LLM subgoals alone, (b) VLM critic alone, and (c) their combination against strong video-policy and hierarchical-planning baselines. Absent these, the claim that the joint framework is necessary and superior remains under-supported.","section":null},{"comment":"Action-conditioning of the video generator is stated as a limitation of prior work that RoboTALES fixes, yet the training objective that makes rollouts reliably action-conditional (and how actions are extracted into the final policy) needs a precise statement—loss terms, conditioning interface, and any distillation or inverse-dynamics step—so that the single-stage claim can be verified and compared to two-stage imagination-then-plan pipelines.","section":null}],"minor_comments":[{"comment":"Define notation for subgoal sequences, VLM reward signals, and the generator conditioning interface early and consistently; the abstract-level description leaves the single-stage training loop underspecified for replication from text alone.","section":null},{"comment":"Clarify whether the VLM critic family overlaps any automatic evaluation judges used on RoboCasa/LIBERO10, and report any reward-threshold or weighting choices as free parameters.","section":null},{"comment":"Add a short limitations discussion on VLM bias on synthetic frames and failure modes when hierarchical decompositions are incorrect.","section":null},{"comment":"Ensure figure captions and method diagrams label the two innovations and the policy-extraction path so readers can map claims to components without relying solely on the abstract.","section":null}],"recommendation":"major_revision","confidential_remarks":"Only the abstract was supplied in the initial packet; my major comments track the load-bearing VLM-alignment and ablation gaps flagged by the stress test and are framed as fixable within a revision if the full manuscript already contains (or can add) calibration and isolation experiments. If the full paper already reports those analyses with solid numbers, the recommendation can move to minor_revision on a quick re-read. Scope fits cs.RO / robot learning venues that accept systems+empirical imagination-for-control work."},"author_rebuttal":{"model":"grok-4.5","summary":"We thank the referee for a careful and constructive review. The three major comments correctly identify places where the manuscript must more rigorously attribute long-horizon gains to the VLM critic, isolate the hierarchical planner from the critic, and state the action-conditioning objective with enough precision to support the single-stage claim. We agree with the substance of all three points and will revise the paper accordingly: adding VLM–progress calibration and reward-control experiments, expanding evaluation tables with seeds/error bars and the requested ablations, and writing an explicit statement of the conditioning interface, losses, and policy extraction path. We believe these revisions will make the empirical contribution clearer and more attributable without changing the core method.","responses":[{"response":"We agree. Attribution of anti-drift and task alignment to the VLM critic is not yet supported by the analyses the referee requests, and long-horizon gains cannot be cleanly credited to innovation (2) without them. In the revision we will (i) report an explicit calibration/correlation study of VLM reward scores against ground-truth success and intermediate progress on held-out trajectories in the same visual domains used for RoboCasa and LIBERO10, and (ii) add the requested control suite that freezes the hierarchical LLM planner and replaces VLM rewards with constant, noisy, and shuffled scores (and, where informative, a no-critic baseline). We will discuss how strongly VLM scores track true progress and how much of the long-horizon improvement remains under degraded reward signals. These additions directly address the attribution concern.","revision_made":"yes","referee_comment":"The central anti-drift and task-alignment claim rests on the VLM critic assigning rewards on synthetic video that track true task progress. The manuscript needs an explicit calibration or correlation analysis (VLM scores vs. ground-truth success/progress on held-out real or sim trajectories in the same visual domain) and a control that freezes the hierarchical planner while replacing VLM rewards with constant, noisy, or shuffled scores. Without this, long-horizon gains on RoboCasa/LIBERO10 cannot be attributed to innovation (2) rather than the LLM planner, video-backbone capacity, or VLM idiosyncrasies."},{"response":"We agree that the current evaluation presentation under-supports the claim that the joint framework is necessary and superior. The revision will include full result tables for RoboCasa and LIBERO10 with all baselines, sample sizes, multiple seeds, and error bars, with long-horizon subsets highlighted. We will also report the three ablations requested—(a) hierarchical LLM subgoals alone, (b) VLM critic alone, and (c) their combination—alongside strong video-policy and hierarchical-planning baselines already used in the paper, so that the contribution of each component and of the joint system is visible. Where a component is already partially present in the experiments, we will reorganize and complete the comparison rather than claim novelty for missing cells. This should make the necessity and superiority claims empirically checkable.","revision_made":"yes","referee_comment":"Headline outperformance is asserted for RoboCasa and LIBERO10, especially long-horizon tasks, but the evaluation must report full tables with baselines, sample sizes, seeds/error bars, and ablations that isolate (a) hierarchical LLM subgoals alone, (b) VLM critic alone, and (c) their combination against strong video-policy and hierarchical-planning baselines. Absent these, the claim that the joint framework is necessary and superior remains under-supported."},{"response":"We agree that the manuscript does not yet state the action-conditioning path with enough precision for the single-stage claim to be verified or fairly compared to two-stage imagination-then-plan pipelines. In the revision we will add a dedicated subsection that specifies: (i) the conditioning interface (how actions and hierarchical subgoals enter the video generator), (ii) the full training objective and loss terms that encourage action-conditional, task-aligned rollouts, including the role of VLM reward feedback, and (iii) exactly how actions are obtained for the final robot policy (e.g., any inverse-dynamics, distillation, or direct readout step) and how this remains a single training stage rather than a separate plan-then-act pipeline. We will also briefly contrast this interface with typical two-stage setups so the single-stage claim is operationally clear. No change to the method is required for this clarification; the text will match the implemented system.","revision_made":"yes","referee_comment":"Action-conditioning of the video generator is stated as a limitation of prior work that RoboTALES fixes, yet the training objective that makes rollouts reliably action-conditional (and how actions are extracted into the final policy) needs a precise statement—loss terms, conditioning interface, and any distillation or inverse-dynamics step—so that the single-stage claim can be verified and compared to two-stage imagination-then-plan pipelines."}],"tokens_in":1766,"tokens_out":1078,"duration_ms":26918,"standing_objections":[]},"desk_editor":{"model":"grok-4.5","letter":"RoboTALES is a systems paper that couples hierarchical LLM subgoal planning and a VLM critic into a video generator so imagined futures stay on task, then trains policies from those futures. That is the whole pitch. If the long-horizon gains on RoboCasa and LIBERO10 hold under real ablations, it is a usable recipe people will actually try.\n\nWhat is new is the single-stage packaging, not the parts. Hierarchical LLM planning, VLM rewards, and video backbones for control are each established. The paper’s contribution is wiring them so the generator is anchored by abstract subgoals and reward feedback, then extracting policies. Framing the failure mode—drift and weak action-conditioning in pretrained video models—is clear and honest. Public code is a real credit.\n\nThe soft spot is load-bearing and matches the stress-test note. Innovation (2) and the anti-drift claim require that VLM scores on synthetic rollouts track true task progress. From what we have, there is no reported calibration of those scores against ground-truth success on real or sim trajectories, no check for bias on synthetic frames, and no ablation that freezes the planner and replaces the VLM with noisy, constant, or shuffled rewards. Without that, the long-horizon wins could come from the hierarchical planner alone, from backbone capacity, or from VLM idiosyncrasies rather than genuine task alignment of the futures. Free parameters (critic thresholds, decomposition granularity, training hypers) are also invisible at abstract level. I am not saying the results are wrong—only that the causal chain from “reasoning-guided imagination” to “better policies” is not yet secured.\n\nThis is for robot-learning people working on long-horizon manipulation and anyone using video generative models as policy backbones. It is a serious systems contribution on its face, not a restatement. Send it to peer review. Ask referees specifically for VLM-reward calibration and planner-only / critic-ablated controls. Do not desk-reject.","headline":"Clean systems recipe for LLM-guided, VLM-scored video futures in robot learning; the anti-drift claim hangs on an unshown VLM–progress link.","tokens_in":2891,"tokens_out":507,"would_cite":false,"duration_ms":22459,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"RoboTALES turns video generators into robot policies by anchoring imagined futures with LLM subgoals and VLM rewards","keywords":["robot learning","video generative models","visuomotor control","hierarchical planning","vision-language models","task-aligned futures","long-horizon manipulation","RoboCasa"],"falsifier":"Train and evaluate the same generator with the VLM critic ablated or replaced by random rewards; if success rates on long-horizon RoboCasa and LIBERO10 tasks collapse to the level of unguided video-generation baselines, the claim that VLM feedback is what keeps futures task-aligned is falsified.","tokens_in":2908,"feed_emoji":"🤖","tokens_out":815,"duration_ms":13789,"temperature":0.7,"pith_summary":"Pretrained video generative models can imagine what a robot might see next, but those imagined futures often wander away from the task and do not reliably tell the robot what action to take. RoboTALES is a single-stage framework that fixes both problems. A hierarchical language model first breaks a complex instruction into ordered subgoals that steer the generator's imagination. A vision-language model then scores the imagined video clips and feeds those scores back as rewards so the generator stays locked on task progress. The resulting task-aligned simulated futures are used to train a robot policy. On the RoboCasa and LIBERO10 manipulation suites the method consistently beats prior approaches, with the largest gains on long-horizon tasks that require multi-step reasoning.","feed_headline":"LLM subgoals and VLM rewards keep robot video futures on task","feed_subtitle":"Policies trained on those futures beat prior methods on long-horizon RoboCasa and LIBERO10 tasks","key_machinery":"The dual-anchor loop: a hierarchical LLM planner that decomposes the task into subgoals and a VLM critic that scores imagined futures with task-progress rewards. Together they keep the video generator's internal representations and rollouts aligned with the intended goal.","core_discovery":"RoboTALES shows that a video generator can be turned into a reliable source of robot policies when it is jointly guided by hierarchical LLM subgoal planning and VLM-based reward feedback. The two signals keep the generator's rollouts temporally consistent and goal-focused, so the actions extracted from them succeed on diverse long-horizon manipulation tasks.","pith_inferences":["The same dual-anchor idea could be tried with smaller, domain-specific language and vision models if full LLM/VLM stacks prove too heavy for onboard robots.","If VLM reward noise is the main remaining failure mode, calibrating the critic on a small set of real robot videos might further close the sim-to-real gap.","Hierarchical subgoal plans may transfer to other generative backbones (diffusion, flow matching) beyond the video models tested here."],"forward_implications":["Robot policies extracted from RoboTALES futures will outperform existing video-generation and imitation baselines on multi-step household manipulation benchmarks.","Long-horizon tasks that previously drifted under pure video imagination become solvable once subgoal guidance and reward feedback are added.","A single-stage training pipeline can replace multi-stage planning-plus-control pipelines that separately invent futures and then learn actions.","Public code and models enable direct reproduction of the gains on RoboCasa and LIBERO10."],"fun_headline_variants":["LLM subgoals and VLM critics align video futures for robot policies","Hierarchical LLM plans plus VLM rewards stop robot video drift","RoboTALES steers generative futures into coherent long-horizon actions","Subgoal sequences and reward feedback extract reliable robot policies","Reasoning-guided video generators yield task-focused robot control"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that a vision-language model's reward scores on synthetic imagined videos are accurate enough and well enough aligned with true task progress to keep the generator goal-focused and to yield policies that work in real evaluation environments.","fun_headline_variants_meta":{"raw":{"variants":["LLM subgoals and VLM critics align video futures for robot policies","Hierarchical LLM plans plus VLM rewards stop robot video drift","RoboTALES steers generative futures into coherent long-horizon actions","Subgoal sequences and reward feedback extract reliable robot policies","Reasoning-guided video generators yield task-focused robot control"]},"model":"grok-4.5","cost_usd":0.00721,"raw_usage":{"total_tokens":1765,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":72100000,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":927,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":89,"duration_ms":10925,"temperature":1.0,"reasoning_tokens":927,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T19:13:15.986658+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the same generator with the VLM critic ablated or replaced by random rewards; if success rates on long-horizon RoboCasa and LIBERO10 tasks collapse to the level of unguided video-generation baselines, the claim that VLM feedback is what keeps futures task-aligned is falsified.","supporting_citations":[],"review_version":1}