{"id":"70e26cc5-3d85-478b-9ac2-c3b1dbad05aa","arxiv_id":"2508.16943","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A staged teacher-student pipeline lets a simulated humanoid relocate two objects in sequence without resets, from egocentric RGB and language, beating the single-task baseline on 350 training and 66 unseen layouts.","lead":"This paper presents a simulated humanoid that can perform two sequential pick-and-place tasks in one uninterrupted episode, guided by camera images and spoken commands. It solves the transition between tasks by training a first controller to step back after placing an object, a second to recover from that pose, then distilling both into a single vision-language-action policy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 72.4% success-all headline is for the oracle teacher, not the distilled RGB+language student; the student evaluation also lacks a specified inference-time transition trigger, so the central reset-free claim is not yet demonstrated.","rationale":"The reader correctly identifies that the transition detection mechanism is under-specified and that the manuscript is not self-consistent, but their strongest claim conflates the oracle teacher results with the distilled student results. The 72.381% and 64.646% Success All figures are for HumanoidVerse-Teacher, which is trained with privileged state; the RGB+language student reaches 40.952% and 40.909% on the same tables. That distinction matters because the central contribution is the student policy. The paper does provide a multi-stage curriculum, ablations, and qualitative comparisons, which are valuable, but the core empirical claim for a reset-free RGB+language policy requires knowing how the student handles the first-to-second-task transition at inference. Algorithm 3's rule-based teacher switching uses privileged simulator signals; without a specified inference-time counterpart, the student's Success All may be achieved with oracle-timed instructions. This is a concrete, load-bearing gap. The title/body mismatch compounds the problem by making the submitted artifact difficult to evaluate as a coherent preprint. The verdict should remain REJECT, although a corrected resubmission with matching metadata, a specified inference-time transition policy, and multi-seed student results could make the claim testable.","tokens_in":11353,"tokens_out":6464,"duration_ms":79241,"concrete_test":"Access the evaluation harness and re-run the student policy on the 350 training tasks with Algorithm 3's privileged trigger disabled: issue the second language instruction only when a vision-only/heuristic detector (or a fixed human protocol) confirms the first object is placed and the robot has retreated, and measure Success All. Also verify that no environment reset occurs between the two object manipulations, and report the teacher and student rows separately in any headline claim. If student Success All falls to near the HumanVLA-S level, the reported result depends on oracle-timed switching; if it remains ~40%, the reset-free claim has some support but still needs multi-seed statistics and released code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a single RGB+language policy performs continuous multi-object rearrangement—rests on Table 1's 'HumanoidVerse-S' row, not the 'HumanoidVerse-T' row. The reader's 72.381% (training) and 64.646% (unseen) numbers are teacher results with privileged object poses/goals; the distilled student achieves 40.952% and 40.909% Success All. That is still above HumanVLA-S (0%), but it is a different, weaker claim. More importantly, the paper never specifies how the student decides when the first task is finished at inference. Algorithm 3 uses oracle object speed, object-to-goal distance, and robot-to-object distance to trigger the second instruction during DAgger training; if the same rule is used during evaluation, the transition is effectively timed by privileged information, and the student's success-all number does not demonstrate perception-driven reset-free chaining. The body also omits whether evaluations use multiple seeds/error bars, the VLA architecture, and the inference-time transition mechanism; the arXiv title/abstract ('LHM-Humanoid') does not match the body ('HumanoidVerse'). Thus the evidence for the load-bearing claim is currently unverifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces HumanoidVerse, a teacher–student framework for vision-language-guided humanoid control that aims to perform continuous, reset-free multi-object rearrangement in a simulated physics environment. A teacher policy is trained with goal-conditioned RL and an AMP style reward through a multi-stage curriculum: single-object rearrangement, release-and-step-back behavior, and second-object manipulation from non-canonical initial states. Two teachers are then distilled via DAgger into a single VLA student that observes only egocentric RGB and natural language. The paper reports experiments in Isaac Gym on 350 training tasks and 66 unseen tasks, comparing against HumanVLA, with success rates and placement distances. The central claim is that this is the first system enabling physically stable, continuous multi-object rearrangement by a humanoid, and that the distilled RGB+language policy substantially outperforms prior single-task methods.","tokens_in":11651,"tokens_out":3504,"duration_ms":46955,"significance":"If the central claim is substantiated, the work would be a meaningful step in physics-based humanoid manipulation: it attacks a genuinely important problem—transition handling in reset-free long-horizon tasks—and introduces a sizable dataset of 350 two-object rearrangement tasks across four room layouts. The multi-stage curriculum, the explicit reward shaping for release and retreat, and the dynamic teacher-switching mechanism during distillation are all reasonable ingredients. However, the current manuscript does not provide verifiable evidence for the strongest form of the claim. The student policy's ability to detect task completion from RGB alone is never specified or evaluated, the VLA architecture is not described, and the empirical comparison is limited to a single baseline with no variance information. The significance is therefore conditional on substantial additional specification and experimentation.","major_comments":[{"comment":"The teacher-switching mechanism uses privileged simulator quantities: object speed, object-to-goal distance, and robot-to-object distance. The student policy receives only egocentric RGB and language, yet the paper never states how, at inference, the student decides that the first object has been successfully placed and that it should wait for or trigger the second instruction. If the same oracle rule is used during evaluation, the reported Success All numbers for HumanoidVerse-S (Tables 1 and 2) do not demonstrate perception-driven reset-free chaining—they are effectively timed by privileged information. This is load-bearing for the central claim and must be resolved by specifying a learned transition detector, or by clearly evaluating with and without oracle switching and reporting both.","section":"§3.4, Algorithm 3"},{"comment":"The headline numbers emphasized in the text (72.381% and 64.646% Success All in the abstract, introduction, and Section 4.2) refer to the teacher policy, while the distilled student achieves 40.952% and 40.909%, respectively. The reset-free claim, however, rests on the student row. The paper conflates these two levels throughout. In addition, the results are point estimates with no error bars, no number of seeds, and no statistical significance testing. The comparison set is also only HumanVLA, despite the abstract claiming comparison against end-to-end RL, hierarchical RL, and prior human-scene-interaction methods. This makes the quantitative claim currently unverifiable.","section":"§4.2, Tables 1 and 2"},{"comment":"The student model is described only as a 'Unified VLA Model'. The architecture is never specified: no backbone, no action representation, no observation preprocessing, no training loss, no DAgger iteration count, and no details on how the natural language instruction is encoded or grounded. Without these details the distillation result cannot be reproduced or assessed for confounding factors such as memory of the first task versus reliance on the second teacher's state. This is a central missing component for a paper whose contribution is a VLA policy.","section":"§3.4, Stage 4"},{"comment":"The success criteria behind Success 1, Success 2, and Success All are not precisely defined. Algorithm 3 refers to 'eval success thresh', and Section 4.1 mentions Dist 1 and Dist 2 but does not state the thresholds used to declare a placement successful, the definition of object-to-goal distance, or how partial credit is handled. Since the ablation study in Table 3 varies conditions that directly affect these thresholds, the reader cannot determine whether the reported differences are meaningful or an artifact of threshold choices.","section":"§4.1, Evaluation metrics"}],"minor_comments":[{"comment":"The arXiv metadata, title, and abstract describe 'LHM-Humanoid', while the body text, figures, and project page refer to 'HumanoidVerse'. The acronym LHM is never defined. This mismatch must be reconciled before any further review.","section":"Title/Abstract vs. Body"},{"comment":"The caption of Figure 6 says 'HumanVLA on the same task as in Figure 6'; it should refer to Figure 5.","section":"Figure 6 caption"},{"comment":"Training epochs per stage are given, but other hyperparameters—AMP reward weight λAMP, reward weights α_i, DAgger aggregation schedule, learning rates—are not provided. Please include a full hyperparameter table.","section":"Appendix A.1, Table 6"},{"comment":"The threshold 'threshobject2goal' is never specified. The same applies to 'speed thresh', 'distance thresh', and 'time thresh' in Algorithm 3. These values are needed to interpret the ablations in Table 3.","section":"§3.2, Algorithm 2"},{"comment":"The qualitative comparisons (Figures 5–13) would be more informative with timestamps or step annotations indicating when the first task ends and the second instruction is given, and with labels for success/failure in the HumanVLA sequence.","section":"Qualitative figures"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a substantially revised or renamed version of itself: the arXiv metadata and abstract describe LHM-Humanoid, while the body is entirely about HumanoidVerse. I would ask the editor to have the authors clarify the submission identity. The central technical problem is that the reset-free student claim is not yet demonstrated: the inference-time transition mechanism is unspecified, and the student-level results are conflated with teacher-level results. These are fixable in principle, but the current draft is not ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real idea buried under a messy submission. What's actually new is the reset-free target: a single physics-based humanoid that does two fetch-carry-place episodes back-to-back without an environment reset, plus the two-stage curriculum that makes it work—a release-and-step-back behavior after the first placement, and a second teacher trained from non-canonical start poses. The dataset, 350 two-object tasks across four room types, is a useful resource. Teacher results are genuinely strong: 72.4% success-all on training tasks versus 5.7% for HumanVLA-Teacher. Student results are more modest but still nonzero: 41% versus 0% for HumanVLA-S. That is a real proof-of-concept.\n\nThe soft spots are significant. Most importantly, the paper never specifies how the student decides, at inference, when the first task is done. Algorithm 3 uses privileged object speed, object-to-goal distance, and robot-to-object distance to trigger the second instruction during DAgger training; the body says nothing about the inference-time trigger. If a human gives the second instruction, that's fine but should be stated. If the oracle rule is used, the student's success-all doesn't demonstrate perception-driven chaining. This is a load-bearing gap, not a cosmetic one.\n\nSecond, the arXiv title and abstract describe LHM-Humanoid and promise comparisons to end-to-end RL and hierarchical RL; the body is HumanoidVerse and compares only to HumanVLA. That mismatch is a submission hygiene failure that would trip any reviewer. Third, there are no error bars or seeds, the VLA architecture is not described, and no code or data are provided. Fourth, the only baseline is HumanVLA; the end-to-end RL and hierarchical RL promised in the metadata are absent.\n\nOn the merits, the central idea is not obviously wrong. The reward design is concrete, the ablations of the two curriculum stages show their contribution, and the problem—continuous multi-object rearrangement in a humanoid—is worth solving. But the evidence for the student's capability is incomplete, and the transition-detection ambiguity means the headline number could be misleading.\n\nWho is this for? Anyone working on humanoid control or teacher-student distillation for locomotion. I would not cite it in its current form; I'd wait for a revised version. A serious referee could help the authors fix the transition-detection specification, add seeds, and align the metadata with the body. This should go to peer review, not be desk-rejected: the problem is real, the approach is a genuine extension, and the flaws are fixable.","headline":"Genuine curriculum and dataset, but the student's reset-free chaining is not yet demonstrated; needs revision, not rejection.","tokens_in":12173,"tokens_out":4255,"would_cite":false,"duration_ms":49875,"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 simulated humanoid can complete consecutive fetch-carry-place commands in one episode by learning to end each cycle in a recoverable state.","keywords":["humanoid control","long-horizon manipulation","multi-object rearrangement","vision-language-action","physics-based simulation","reinforcement learning","DAgger distillation","release-and-retreat"],"falsifier":"Run the trained student on two-object episodes where the first object is placed but still sliding, or where the goal is partially occluded, and compare the timestep at which the student begins the second task with the oracle rule in Algorithm 3. If the student starts the second instruction before the first object is effectively at rest, its both-task success should collapse, showing that its RGB-only transition detection is the true bottleneck.","tokens_in":11244,"feed_emoji":"🤖","tokens_out":10857,"duration_ms":113453,"temperature":0.7,"pith_summary":"The paper introduces HumanoidVerse, a system that tries to show a physically simulated humanoid can work continuously without resets: hear one instruction, walk to an object, carry it to a goal, release it, step back, wait, then hear a second instruction and repeat, all in a single episode. The paper identifies the transition between tasks, not the individual motions, as the real obstacle, because placing an object leaves the character off-balance in a pose no single controller has seen before. It answers this by learning a release-and-retreat ending that brings the character into a recoverable state, then handing off to a second controller trained from exactly those states. The paper backs the claim with 350 two-object rearrangement tasks across four room layouts and reports that the two-step chain succeeds far more often than prior single-object controllers, on both training and unseen scenes.","feed_headline":"One humanoid policy completes two object moves without a reset","feed_subtitle":"In simulation the teacher completes both moves in 72% of layouts; the vision-language student keeps 41%.","key_machinery":"The load-bearing mechanism is a multi-stage curriculum with two teacher controllers and a rule-based handoff. Stage 1 pretrains a single-object rearrangement policy with an adversarial motion prior; Stage 2 adds a release-and-step-back reward so that after placing the first object the robot leaves it undisturbed and stops about a meter away; Stage 3 trains a second teacher to take over from the resulting non-canonical postures; Stage 4 distills both teachers into one vision-language-action policy using DAgger with a switching rule based on object speed, object-to-goal distance, and robot-to-object distance. The 'recoverable region' is the abstract target: the terminal state of the first cycl","core_discovery":"The paper's central claim is that the hard part of long-horizon humanoid manipulation is the handoff between tasks, not the individual fetch–carry–place motions. After a humanoid places the first object it is left in a non-canonical, off-balance pose, and naive end-to-end reinforcement learning cannot reliably start the next task from such a state. The paper's solution is to train the first controller to end each placement with a release-and-retreat behavior that steers the terminal state into a recoverable region: close enough to have disengaged without disturbing the placed object, and far enough to leave a balanced starting point for the next cycle. A second controller is then trained fro","pith_inferences":["Editorial inference: the same curriculum should extend beyond two objects if every release-and-retreat leaves the character in a recoverable region; nothing in the method is tied to the number two.","Editorial inference: the oracle rule in Algorithm 3 could be replaced by a learned transition detector trained on egocentric RGB, which would directly test whether the student's visual perception can carry the reset-free chain without privileged simulator signals.","Editorial inference: the release-and-retreat reward could be reused as a standalone safety primitive for real humanoid hardware, since stepping back after placing an object is a simple way to avoid knocking the just-placed object during the next command."],"forward_implications":["The near-equality of Success All (72.4%) and Success 2 (72.9%) on training tasks means the second cycle can be completed without disturbing the first placement, which is exactly what reset-free chaining requires.","The ablations show the release-and-retreat stage and the from-arbitrary-posture stage are both necessary: dropping Stage 2 lowers both-task success from 72.4% to 56.2%, and dropping Stage 3 lowers it to 35.0%.","The rule-based switching conditions each carry weight: removing the object-speed check or the object-to-goal distance check reduces both-task success by roughly 17 to 19 percentage points, so the transition signal is an active component of the pipeline, not a detail.","Because the four-stage curriculum is distilled into a single RGB-plus-language policy, the final system does not need privileged object poses at execution time; the student runs the whole chain with the same observations a person would have."],"supporting_citations":[{"why":"Supplies the AMP adversarial motion prior used as the style reward in the teacher stages, anchoring behavior to human-like motion.","marker":"Peng et al. (2021)"},{"why":"Supplies DAgger, the imitation-learning algorithm used to distill the two teacher policies into the single vision-language student.","marker":"Ross et al. (2011)"},{"why":"HumanVLA is the prior single-object vision-language humanoid system this work extends; it provides the pretrained checkpoint, the HITR dataset basis, and the main baseline.","marker":"Xu et al. (2024)"},{"why":"Isaac Gym is the GPU simulator in which all 350 training tasks and 66 unseen tasks are trained and evaluated.","marker":"Makoviychuk et al. (2021)"}],"fun_headline_variants":["Simulated humanoid carries objects continuously, solving tricky handoffs","No resets: one policy keeps moving objects across scenes","Long-horizon humanoid control: continuous transport without resets","Reset-free humanoid transport: the trick is knowing when to let go"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes the student can detect from its camera when the first object is safely placed and when to stop and wait, because the training-time teacher switch in Algorithm 3 uses simulator state—object speed, object-to-goal distance, and robot-to-object distance—that the camera never sees.","fun_headline_variants_meta":{"raw":{"variants":["Simulated humanoid carries objects continuously, solving tricky handoffs","No resets: one policy keeps moving objects across scenes","Long-horizon humanoid control: continuous transport without resets","Reset-free humanoid transport: the trick is knowing when to let go"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001272,"raw_usage":{"total_tokens":5112,"prompt_tokens":885,"completion_tokens":4227,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":4154}},"tokens_in":629,"tokens_out":4227,"duration_ms":30253,"temperature":1.0,"reasoning_tokens":4154,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:07:06.299847+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained student on two-object episodes where the first object is placed but still sliding, or where the goal is partially occluded, and compare the timestep at which the student begins the second task with the oracle rule in Algorithm 3. If the student starts the second instruction before the first object is effectively at rest, its both-task success should collapse, showing that its RGB-only transition detection is the true bottleneck.","supporting_citations":[],"review_version":1}