{"id":"359194eb-4e27-40c7-94c1-e00db1a35f6c","arxiv_id":"2607.18016","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Persistent role-indexed 3D object tokens that condition both action generation and geometric verification improved a GR00T-N1.7 humanoid's loco-manipulation success from 39/80 to 71/80 across eight real-world task families.","lead":"This paper describes a robot-system upgrade for humanoid machines: the robot keeps a running 3D memory of task objects—cups, baskets, handles—and both plans its motions and checks its own work against that memory. On a Unitree G1 humanoid, the system succeeded in 71 of 80 physical trials where the matched baseline succeeded in 39, and the gain was biggest on tasks that require tracking objects across several steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main comparison confounds persistent object tokens with extra checkpoint fine-tuning: POT variants use an object-token fine-tuned checkpoint; direct/verifier-only use an un-fine-tuned checkpoint (§4.3). Gains may reflect additional adaptation, not the token abstraction.","rationale":"I agree with the reader's weakest assumption. The paper's central contribution is the shared object loop: the same persistent 3D object records condition action generation and verification. The strongest claim asserts that this object-state loop is the causal driver of the large real-world improvement. But §4.3 reveals that the POT-token variants use a different checkpoint — one fine-tuned with demonstrations paired with object-token sidecars — while the direct and verifier-only variants do not. This is a textbook confound: the comparison varies both the inference-time conditioning and the training procedure. The ablation 'POT tokens only' is meant to isolate token conditioning, but it is compared against an un-fine-tuned direct checkpoint, so its 16-point gain over direct baseline may reflect additional adaptation rather than the token representation itself. The verifier-only variant is less confounded (it adds only runtime checks on the direct checkpoint), but the main result and the token-only ablation are the key supports for the central attribution. No code, artifacts, error bars, or significance tests are provided to mitigate this. Thus the verdict should remain CONDITIONAL pending the fine-tuned-without-tokens control, which would directly separate the effect of persistent object tokens from the effect of additional fine-tuning. The external Being-0 comparison is weaker evidence but not load-bearing for the central claim; the checkpoint confound is the correct focal point.","tokens_in":10190,"tokens_out":3001,"duration_ms":34087,"concrete_test":"Train a control with identical demonstrations, fine-tuning budget, and checkpoint-selection rule, but with object-token sidecars replaced by constant padding/masked slots during training and at inference (so no token information is available). Evaluate the same 40-trial ablation subset and, if possible, all eight Table 1A tasks. If this control reaches or approaches the POT-token-only success (e.g., ≥26/40 or comparable total), then the reported token gains are not attributable to persistent object tokens. If it stays near direct baseline (≈15/40), the fine-tuning confound is largely refuted and the central attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"§4.3 states: 'the direct and verifier-only variants use the same direct action-expert checkpoint, while the POT-token variants use the same object-token fine-tuned checkpoint, demonstration sidecars, fine-tuning budget, and checkpoint-selection rule.' Because the object-token branch adds learned parameters and the fine-tuning procedure updates action-head parameters on demonstrations paired with token sidecars, POT-token execution differs from the direct path by two coupled factors: (1) the presence of persistent 3D object tokens at inference, and (2) a different, further-trained checkpoint. The headline matched comparison (39/80 → 71/80, Table 1A) therefore does not isolate the shared object-state loop, and the ablation 'POT tokens only' (31/40 vs 15/40, Table 2A) attributes to token conditioning what could partly come from extra training updates, sidecar-paired data distribution, or the new projection parameters. Since the strongest claim is causal — persistent object-centered state improves verifiable humanoid VLA execution — the causal attribution is not yet cleanly established. A fine-tuned-without-tokens control is needed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Persistent Object Tokenization (POT), a role-indexed 3D object memory that is refreshed from RGB-D observations and serialized into fixed-slot tokens for a GR00T-N1.7 whole-body action head. The same memory is also read by a geometric predicate supervisor that verifies subtask completion and triggers recovery. The system, POT-VLA, is evaluated on a Unitree G1 across eight real-world task families, reporting 71/80 successes versus 39/80 for a direct baseline, and 44/50 on a Being-0-aligned service-task reference versus 37/50 reported by the Being-0 paper. The authors conclude that persistent object-centered state reduces object-state divergence and enables verifiable humanoid loco-manipulation.","tokens_in":10508,"tokens_out":3755,"duration_ms":43738,"significance":"If the causal claim is correct, the paper makes a useful contribution: it shows concrete gains from making object state explicit, persistent, and shared between action generation and verification in a real humanoid system. The strengths are the real-robot evaluation, the use of a fixed action backbone and embodiment across conditions, the separate ablations of token conditioning and verification, and the generalization tests under object-state shifts. However, the main comparison is confounded because the POT variants use an additional object-token fine-tuned checkpoint while the direct baseline does not. The stated success criterion is also the same geometric predicate set used by the verifier to gate completion and trigger recovery. These issues mean the reported improvements are plausible but not yet cleanly attributable to the persistent object-token abstraction. The paper's central technical idea is defensible, but the experimental design needs strengthening before the headline claim can be accepted.","major_comments":[{"comment":"The headline comparison is not a matched control. The direct and verifier-only variants use the direct action-expert checkpoint, while the POT-token variants use the object-token fine-tuned checkpoint, demonstration sidecars, fine-tuning budget, and checkpoint-selection rule. Thus the 39/80 → 71/80 improvement includes at least two coupled factors: the presence of persistent object tokens and an additional fine-tuning stage with sidecar-paired data and new projection parameters. The claim that POT-VLA improves a 'matched direct baseline' is therefore not supported. Please add a control that fine-tunes the direct action expert with the same demonstrations and budget but does not receive object tokens at inference, or otherwise equalizes the amount of fine-tuning across conditions.","section":"§4.3, Table 1A"},{"comment":"Trial success is defined by whether the final geometric predicates remain satisfied over a temporal window, and those same predicates are used by the supervisor to gate completion and trigger recovery. For POT-VLA, a failed predicate can lead to retry/recovery before success is recorded; the direct baseline is evaluated without this feedback. The metric is therefore aligned with the mechanism being added, which can inflate the apparent gain even if the action policy itself is unchanged in quality. Please report first-attempt success (before recovery), the number of recovery attempts per trial, and an independent human or externally defined verification of final states to separate improved action generation from improved recovery gating.","section":"§4.1 and §3.3"},{"comment":"The ablation subset is explicitly chosen as 'four tasks with large state-divergence gaps in Table 1.' Selecting tasks based on the outcome variable biases the estimated contribution of token conditioning upward. The claim that 'POT tokens provide the largest gain' (15/40 to 31/40) is based on this non-random subset. Please run the ablation on the full eight-task suite or pre-register the subset selection before seeing Table 1 results. The small per-setting sample sizes (10 trials) also warrant confidence intervals or a simple significance test.","section":"§4.3, Table 2A"}],"minor_comments":[{"comment":"The text contains an unresolved 'Table??' reference. It should be replaced with the actual table number.","section":"§4.1"},{"comment":"The sentence 'The matched comparison supports the state-divergence hypothesis' overstates the evidence: the experiment does not directly measure object-state divergence, and the comparison is confounded as noted above.","section":"§4.2"},{"comment":"Per-subtask predicate thresholds ν, stability windows n, retry budgets, timeouts, and chunk horizons are hand-specified. A sensitivity analysis over these values would help establish that the reported gains are not driven by favorable calibration.","section":"§3.1, §3.3"},{"comment":"The fine-tuning details are underspecified: demonstration count, number of gradient steps, learning rate, and checkpoint-selection criterion are not reported. This is needed for reproducibility and for judging whether the fine-tuned control would be fair.","section":"§4.3"},{"comment":"The Being-0 comparison uses paper-reported numbers on 'aligned' tasks rather than a local reproduction. This is acknowledged, but the alignment procedure should be described in more detail, and the comparison should not be presented as evidence of equivalence to a matched baseline.","section":"§4.2, Table 1B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is worth pursuing, but the central causal claim currently rests on a confounded comparison and a success criterion that overlaps with the verification mechanism. I would ask the authors to add the fine-tuned-without-tokens control and to report first-attempt versus post-recovery success before considering acceptance. The external Being-0 comparison is not critical to the contribution and should be de-emphasized or made clearly secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate systems paper with a clear problem, a plausible abstraction, and real hardware results. The main claim — that shared role-indexed 3D object tokens improve humanoid VLA execution and verification — is not yet cleanly established, because the comparison conflates the token loop with extra fine-tuning. But the confound is disclosed, the ablation acknowledges it, and the fix is straightforward. Worth refereeing.\n\nWhat's new: the specific integration of persistent role-indexed 3D object tokens that are consumed by a whole-body action head (GR00T-N1.7) and reused by a geometric predicate verifier to gate completion and trigger recovery. Components exist, but this particular coupling — same refreshed object memory driving both action and verification — is a real contribution. The 80-trial real-robot comparison is new. The problem framing as 'object-state divergence' is useful, even if it is a reframing of closed-loop control concerns.\n\nWhat's good: the system design is coherent, the token schema is concrete, the ablation removes one part at a time (verifier only, tokens only), and the paper is transparent about the checkpoint difference in Section 4.3. They also state the Being-0 reference is not a local reproduction. That's honest.\n\nSoft spots: the headline 39/80 to 71/80 compares POT-VLA, which was fine-tuned with object-token sidecars, against a direct baseline that was not fine-tuned. That means the gain could be from extra adaptation rather than the token abstraction. The 'POT tokens only' ablation has the same issue. A fine-tuned-without-tokens control would settle it. Also: ablation tasks were chosen for large expected gaps, no error bars or significance tests, per-subtask thresholds and retry budgets are hand-calibrated, and no artifacts are released. The circularity point — success is defined by the same predicates the verifier uses — is real but somewhat inherent: a verifier that checks the goal is partly the point. Still, it means the success metric is not independent of the method.\n\nThe Being-0 comparison should be downgraded in the text: it's a copied number, not a benchmark.\n\nWho this is for: people working on humanoid VLA, object-centric grounding, and closed-loop execution. A serious referee could help tighten the evaluation. My recommendation: send to peer review, require the fine-tuned control and per-task variance, and ask for code/data or a clear release plan.","headline":"Real integration with a real robot and a transparent but unresolved confound: the headline gain may be extra fine-tuning, not the shared object tokens.","tokens_in":11028,"tokens_out":1546,"would_cite":true,"duration_ms":16684,"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":"The paper argues that object-state divergence (the gap between the state used to act and the state used to verify) is the core failure in long-horizon humanoid manipulation, and that one shared, refreshed 3D object memory closes it.","keywords":["humanoid loco-manipulation","vision-language-action policy","object-centric grounding","3D object tokens","execution verification","geometric predicates","closed-loop control","object-state divergence"],"falsifier":"Run the four-task ablation subset with the object-token fine-tuned checkpoint but with object tokens masked out throughout execution (while keeping predicate verification disabled); if success stays near the full POT-VLA level rather than falling toward the direct baseline, the gain is carried by the fine-tuned checkpoint rather than by the token signal.","tokens_in":10050,"feed_emoji":"🤖","tokens_out":7624,"duration_ms":63359,"temperature":0.7,"pith_summary":"The paper tries to establish that the central problem in long-horizon humanoid loco-manipulation is object-state divergence: the state used to condition an action can drift from the state used to verify it, causing false completions and failed recoveries. It proposes Persistent Object Tokenization (POT), a role-indexed 3D object memory built from RGB-D observations, serialized into tokens for a whole-body vision-language-action policy, and reused after each chunk to run geometric predicate checks. On a real humanoid across eight task families, this closed loop raises task success from 39/80 to 71/80 against a matched direct baseline, with the largest gains on tasks that require maintaining 3D relations across multiple action chunks. The broader claim is that persistent, object-centered, verifiable state is a useful abstraction for general humanoid VLA execution, without needing a simulator or learned dynamics model.","feed_headline":"Persistent object tokens lift humanoid success 39/80 to 71/80","feed_subtitle":"The same 3D records guide action and verify it, closing the loop.","key_machinery":"The central mechanism is the role-indexed object memory M, which binds task entities to persistent role slots (TARGET, DESTINATION, SUPPORT, HANDOVER_PARTNER) and refreshes their metric 3D evidence from RGB-D after each executed chunk. This memory is serialized into Persistent 3D Object Tokens: K=8 slots by F=33 features, projected through a LayerNorm→Linear→GELU→Dropout→Linear branch and inserted into the action head's self-attention sequence as state, object, and action tokens. The same memory feeds the geometric predicate supervisor, whose predicates p=(kappa, alpha, op, nu, n) (type, grounded arguments, comparison, threshold, stability window) decide in_progress, done, blocked, failed, o","core_discovery":"POT-VLA conditions a whole-body action expert on Persistent 3D Object Tokens—fixed-slot tensors carrying semantic role, grounding confidence, 3D centroid and extent, and relational features for up to eight task entities—and then verifies the same refreshed object memory with geometric predicates such as containment, support, proximity, displacement, and handover distance. Because action generation and verification draw on identical object records, a subtask advances only when the intended metric relation actually holds in refreshed 3D state. Across eight real-world task families on a real humanoid robot, POT-VLA completes 71/80 trials versus 39/80 for a matched direct baseline; the largest g","pith_inferences":["A clean testable extension would isolate the token signal from the fine-tuning effect: run the object-token fine-tuned checkpoint with the object tokens zeroed or masked, and compare with the same checkpoint using live tokens; if the gap persists, the improvement comes from the token content rather than from checkpoint adaptation.","If persistent role-indexed memory is the operative abstraction, similar gains should appear when the same object-token schema is attached to a different whole-body action expert, provided the action head supports token insertion—an explicit cross-backbone transfer test the paper does not run.","The predicate supervisor's hand-specified thresholds and retry budgets are currently calibrated per subtask; one could infer that auto-tuning these from measured success or variance would make the system less dependent on manual calibration and more portable across robots and scenes.","A further implication is that the same refreshed object memory could power higher-level failure explanations or a learned residual correction model, since it already localizes failures as missed grasp, outside-goal placement, unstable support, or excessive handover distance."],"forward_implications":["If the central claim holds, VLA policies for long-horizon manipulation should expose task-object state explicitly rather than leaving it implicit inside visual-language features.","Sharing one refreshed object memory between action generation and verification should reduce false task completions, since a subtask finishes only when the intended 3D relation is measured, not merely attempted.","The approach should generalize without a simulator or physics engine: POT is built from measurement-grounded RGB-D records, so it applies wherever reliable depth, masks, and calibration are available.","The largest gains on multi-chunk relational tasks suggest that persistent object tokens help most where objects move, become occluded, or change support through contact and transport.","Recovery becomes less expensive: local re-observation, regrounding, and retry can fix object-state errors before any full language-level replan is needed."],"fun_headline_variants":["Same 3D tokens drive and verify humanoid actions, lifting success to 71/80","Persistent 3D object tokens close the loop: 71/80 success","Closed-loop VLA: same 3D records for action and verification, 71/80","Object tokens as ground truth: humanoid loco-manip hits 71/80","Verified object state boosts humanoid VLA to 71/80 (from 39/80)"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation assumes the direct baseline is a matched control that isolates the object-token loop, but the compared variants differ in which checkpoint they use (object-token fine-tuned versus direct) and in hand-specified verification thresholds, retry budgets, and timeouts, so the measured gains cannot be cleanly attributed to persistent 3D object tokens alone.","fun_headline_variants_meta":{"raw":{"variants":["Same 3D tokens drive and verify humanoid actions, lifting success to 71/80","Persistent 3D object tokens close the loop: 71/80 success","Closed-loop VLA: same 3D records for action and verification, 71/80","Object tokens as ground truth: humanoid loco-manip hits 71/80","Verified object state boosts humanoid VLA to 71/80 (from 39/80)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000823,"raw_usage":{"total_tokens":3459,"prompt_tokens":787,"completion_tokens":2672,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":2556}},"tokens_in":531,"tokens_out":2672,"duration_ms":17048,"temperature":1.0,"reasoning_tokens":2556,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:21:33.235616+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the four-task ablation subset with the object-token fine-tuned checkpoint but with object tokens masked out throughout execution (while keeping predicate verification disabled); if success stays near the full POT-VLA level rather than falling toward the direct baseline, the gain is carried by the fine-tuned checkpoint rather than by the token signal.","supporting_citations":[],"review_version":1}