{"id":"2faa82fd-0cf7-4b73-80bd-95dab0c77450","arxiv_id":"2607.14739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FoMoVLA improves VLA robot policies by jointly training future-feature foresight and sparse point tracking, coupled through future-conditioned cross-attention, with auxiliary branches removed at inference.","lead":"This paper proposes FoMoVLA, a training scheme that adds two auxiliary supervision tasks to vision-language-action robot models: predicting future visual features and tracking sparse 2D points, coupled by a cross-attention module. On LIBERO, RoboCasa, and LIBERO-Plus benchmarks it reports higher task success than its base model, at the cost of only about 10 ms extra inference latency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: the final model keeps 16 <Foresight> tokens and a reordered attention mask at inference; ablations vs 'Base Backbone' thus change architecture, not just auxiliary losses, so the reported gains may not be caused by the training-only supervision.","rationale":"The paper proposes a plausible framework and the internal ablations show monotonic improvements, which is encouraging. However, the headline claim of state-of-the-art performance rests on a comparison that conflates the auxiliary training objectives with architectural changes at inference: the final model retains extra <Foresight> tokens and uses a reordered attention mask, while the base backbone does not. Without a control that adds the same tokens/ordering but omits the auxiliary losses, the reported gains (LIBERO 96.5→98.8, RoboCasa 47.8→56.9) cannot be attributed to the proposed supervision. This is a standard experimental-design gap, not an indication of dishonesty. The reader's concern about CoTracker pseudo-label quality is valid but secondary: even if the teacher's tracks are imperfect, a positive controlled result would still show the framework helps; conversely, the missing control could explain the gains regardless of tracking quality. The proposed test is cheap, uses the paper's own infrastructure, and would settle the causal claim. Verdict remains CONDITIONAL, with the explicit condition that the control experiment be run and reported.","tokens_in":15604,"tokens_out":5688,"duration_ms":59101,"concrete_test":"Run a controlled ablation on LIBERO (and ideally RoboCasa): start from the same StarVLA-GR00T base, apply the identical text-first reordering and append K=16 randomly initialized <Foresight> tokens with the same attention mask, but train with only L_action (no L_foresight, no L_track, no FCCA) under the same 30K steps / batch size / eval protocol. Report per-suite and average success; if the control ≥98.8% (LIBERO) / ≥56.9% (RoboCasa), the auxiliary losses are not the cause; if it stays near 96.5%/47.8%, the central claim is supported. Run 3 seeds to also establish error bars.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that training-only auxiliary supervision (future feature prediction + point tracking) improves the VLA policy. But the comparison against the 'Base Backbone' in Table 1 does not isolate these losses. Section 3.3 says the K=16 <Foresight> tokens are retained at inference, and Supplementary A.2 describes a reordered attention mask (Instruction→Image→Foresight→Action, with image tokens no longer before text). So the full model differs from the base in at least three ways: (1) the two auxiliary losses, (2) 16 extra learnable tokens increasing input length and model capacity, and (3) a different token order/attention pattern. The '+ Future Prediction' row includes changes (2) and (3) plus the foresight loss; '+ Tracking' likely includes (3) but not (2). No ablation controls for (2) or (3) alone. For example, a model with the same reordering and 16 randomly initialized extra tokens, trained with only L_action, has not been run. If such a control already reaches ~98% on LIBERO, the paper's attribution of the gains to joint foresight/tracking supervision collapses. The paper's own limitation statement discusses 2D vs 3D motion but never mentions this architectural confound, and the reader's weaker assumption about CoTracker labels, while reasonable, would not explain a gain caused by extra tokens.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FoMoVLA augments a VLA policy with two auxiliary training objectives: compact future-feature prediction via K learnable <Foresight> tokens, and sparse 2D point tracking supervised by a frozen CoTracker-v3 teacher, coupled by a future-conditioned cross-attention module. At deployment the auxiliary heads are discarded, but the <Foresight> tokens and a reordered attention mask are retained. The paper reports state-of-the-art LIBERO and RoboCasa GR-1 Tabletop results, plus LIBERO-Plus zero-shot gains, with ablations attributing the improvement to the two auxiliary losses and their coupling.","tokens_in":15987,"tokens_out":6645,"duration_ms":68305,"significance":"The idea of coupling a global future-state representation with sparse per-point motion cues is well motivated, and the method is specified concretely with explicit loss equations, parameter counts, and component-wise ablations. The reported inference overhead is small and the main success metrics are independent of the auxiliary losses, so there is no deep circularity in the headline result. However, the central attribution of the gains to the auxiliary supervision is not yet established because the comparison against the base backbone changes architecture and input sequence, not only losses; statistical support for the small LIBERO margins is also missing. If these issues are resolved, the paper would be a solid contribution to VLA training with predictive and geometric supervision.","major_comments":[{"comment":"The central comparison does not isolate the auxiliary losses. As described in Sec. 3.3 and Supp. A.2, the full model retains K=16 <Foresight> tokens at inference and uses a reordered attention mask (Instruction->Image->Foresight->Action), while the 'Base Backbone' row in Table 1 presumably uses neither. Thus '+ Future Prediction' adds the foresight loss plus 16 extra learnable tokens plus a new attention topology, and '+ Tracking' adds the tracking loss plus the reordering. No ablation controls for the extra tokens or the attention reorder alone. Please run the missing controls, e.g., base + reordered mask without extra tokens, base + 16 randomly initialized learnable tokens with no auxiliary losses, and the full architecture with auxiliary losses disabled. Without these, the paper's claim that training-only supervision causes the gains is not supported.","section":"Sec. 3.3, 3.5, Supp. A.2, Table 1"},{"comment":"Statistical support is missing for the LIBERO claim. With 20 rollouts per task and 10 tasks per suite, per-suite success rates have sizable uncertainty, and many reported differences are small (e.g., Spatial 98.4 vs 97.8; Object 99.6 vs 98.8; the full-model average 98.8 vs 98.5 for several baselines). No confidence intervals, standard errors, significance tests, or multi-seed results are reported for Tables 1-3. The RoboCasa gain (+9.1 pp) is large, but the LIBERO state-of-the-art claim needs explicit error bars or tests to show that the margins are not noise.","section":"Table 1, Sec. 4.2"},{"comment":"The tracking evaluation in the appendix is circular. The 'ground truth' for ATE and Survival@10px is the frozen CoTracker-v3 teacher, which is the same teacher used to generate the training targets in Eq. (5)-(7). The reported improvement therefore demonstrates that FCCA makes predicted tracks agree more closely with the teacher's pseudo-labels, not that tracking is more accurate in an independent sense. This does not affect the action-success benchmarks, but the claim that FCCA improves point-tracking accuracy should be reframed as 'agreement with the teacher' or evaluated on independently labeled tracks.","section":"Supp. B.2, Fig. 6"}],"minor_comments":[{"comment":"Typo: 'bythemainbythemain' appears in the paragraph on 'Shared image tokens'. The conclusion also begins with a lowercase fragment ('jointly learning...').","section":"Sec. 4.3"},{"comment":"The text says the auxiliary branches incur 'zero parameter overhead', which is technically true, but inference still processes 16 extra <Foresight> tokens and uses a changed attention mask. The phrase 'training-only' should be qualified as 'training-only losses' to avoid confusion with the retained architectural changes.","section":"Sec. 3.5, Table 8"},{"comment":"'Consistently improves' is too strong: for StarVLA-π the FoMoVLA variant decreases on Spatial (-0.6) and Object (-1.0), with the average gain driven by LIBERO-Long (+8.2). Please qualify the scalability claim and discuss the per-suite trade-offs.","section":"Sec. 4.4, Table 5"},{"comment":"The qualitative claim that predicted foresight 'closely reconstructs' ground-truth future features is not backed by a quantitative metric. Reporting the foresight loss or cosine similarity on held-out data would make this claim verifiable.","section":"Sec. 4.5, Fig. 4"},{"comment":"The attention-mask design is important for interpreting the results and should be summarized in the main paper, not only in the supplementary material.","section":"Supp. A.2"}],"recommendation":"major_revision","confidential_remarks":"The architectural confound is the key issue: the paper's central claim that training-only auxiliary supervision improves the policy is not yet supported. If the requested control experiments confirm the result, the contribution would be valuable. The manuscript also relies heavily on 2026 arXiv preprints, including some self-citations; I would encourage the editor to check that the comparison baselines are public and that the implementation details are sufficient for reproduction. No code release is mentioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a coherent, well-written method paper with a genuinely new way to couple visual foresight and point tracking as training-only supervision. The gains over the base backbone are consistent across two benchmarks and three action heads, and the design ablations (grid density, foresight bottleneck, coupling mechanism) are thoughtful. If the effect is real, it's practically useful for VLA training with negligible inference cost. I'd give it credit for that.\n\nThe soft spots are real, though. The stress-test note lands: the full model differs from the Base Backbone in at least three ways—the two auxiliary losses, 16 extra learnable <Foresight> tokens at inference, and a reordered attention mask (instruction before image). The ablations don't isolate these. A control with the same token order and 16 random extra tokens, trained with only the action loss, is missing. Since reordering alone can improve VLA performance in some setups, you can't attribute the full LIBERO jump (96.5 to 98.8) to the auxiliary supervision. The paper's own limitation statement mentions 2D vs 3D motion but never acknowledges this confound, which is a gap.\n\nOther issues: no confidence intervals or significance tests, and many LIBERO margins are within a couple of points. Published baselines are uncontrolled (different backbones, pretraining, data), so the 'state-of-the-art' phrasing is too strong. The tracking-quality evaluation in Figure 6 uses CoTracker as the ground-truth for trajectories that were themselves supervised by CoTracker—so it measures imitation of the teacher, not action-relevant accuracy. That's a minor point, though, since the main result is task success.\n\nWho's this for? People working on VLA training-time regularizers or predictive auxiliary losses will want to read it. It deserves a serious referee, but the referee should insist on the missing control and proper error bars before the central claim is accepted.\n\nMy advice: send it to peer review, but expect revision. The recipe is promising; the evidence as presented doesn't yet support the strongest claims.","headline":"Useful VLA training-time auxiliary supervision with a plausible recipe, but the main ablation is confounded by architectural changes (extra tokens, reordered attention), so the headline gains aren't cleanly attributable to the losses.","tokens_in":16468,"tokens_out":2300,"would_cite":true,"duration_ms":24164,"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":"FoMoVLA claims that training a VLA policy to predict future visual features and sparse 2D point trajectories, coupled through cross-attention, improves manipulation success and zero-shot robustness while adding no inference overhead.","keywords":["vision-language-action models","visual foresight","future feature prediction","point tracking","motion guidance","robot manipulation","auxiliary supervision","LIBERO"],"falsifier":"Train FoMoVLA with the point-tracking supervision replaced by scrambled (randomly permuted or inverted) trajectories from the same teacher, keeping all else equal; if LIBERO-Long and RoboCasa success still rise, the motion signal is not what carries the gain.","tokens_in":15514,"feed_emoji":"🤖","tokens_out":5167,"duration_ms":45057,"temperature":0.7,"pith_summary":"FoMoVLA is a training-only add-on for vision-language-action (VLA) robot policies. The paper argues that reactive policies fail because they map pixels to actions without anticipating what the scene will look like or how objects will move. It proposes teaching the policy two complementary predictions as auxiliary losses: a compact future-feature token that encodes the goal scene, and sparse 2D point trajectories that encode the motion path to that scene. These are coupled by a cross-attention module so motion prediction is conditioned on the predicted future. On LIBERO and RoboCasa benchmarks the full recipe raises average success from 96.5% to 98.8% and from 47.8% to 56.9% respectively, and all auxiliary branches are discarded at deployment.","feed_headline":"Foresight plus point tracking lifts robot policies to 98.8%","feed_subtitle":"Two training-only objectives push LIBERO success from 96.5% to 98.8% with zero extra inference cost.","key_machinery":"The key machinery is the future-conditioned cross-attention (FCCA) module. FCCA takes the hidden states of K learnable <Foresight> tokens (trained via an EMA teacher to reconstruct the final-frame visual features through a compact MAE decoder) and uses them as keys/values in multi-head attention over the spatial image-token hidden states that feed the point-tracker heads. A zero-initialized output projection makes the module an identity at the start of training, so it injects future-state information into motion prediction gradually. The point-tracking branch itself predicts per-frame 2D displacements and visibility for a sparse 8×8 grid of points, supervised by a frozen point-tracker teache","core_discovery":"The central claim is that future feature prediction and sparse point tracking are naturally complementary for embodied control: the former says where the scene is heading, the latter says how points in the scene get there, and jointly supervising a VLA backbone with both, tied together by a future-conditioned cross-attention module, produces representations that generate better action chunks. The paper shows that each objective alone improves the base backbone, that combining them yields further gains, and that coupling them through FCCA is what unlocks the largest long-horizon gains. Because the auxiliary branches are training-only, the deployed model is exactly the base backbone plus a sma","pith_inferences":["A natural extension is to replace the 2D point-tracker teacher with 3D trajectory supervision (e.g., from depth or a 3D tracker), which would directly address the paper's stated limitation that camera-viewpoint and robot-state perturbations remain the weakest gains.","The quantitative tracking metrics reported in the appendix (ATE, Median TE, Survival@10px) could serve as a cheap, rollout-free proxy for tuning foresight quality during training.","The 'goal state plus motion path' decomposition is a general recipe; it could be ported to navigation, dexterous manipulation, or any policy where the bottleneck is anticipating scene change rather than modeling the actuator.","The method's reliance on a frozen teacher's pseudo-labels means the ceiling of the approach is bounded by the quality of that teacher; upgrading the teacher should directly raise the policy's ceiling, and using a weaker teacher should cap it."],"forward_implications":["If FoMoVLA is right, future-state prediction and motion prediction should be trained together, not as separate auxiliary tasks, because the coupling through FCCA is what produces the largest gains on long-horizon tasks.","Since all auxiliary branches are discarded at inference, the recipe adds essentially no deployment cost: only 9.4 ms median latency and 0.1 GB memory for the K extra foresight tokens.","The gains transfer across action-head formulations (flow matching, implicit policy heads), so the supervision recipe is not tied to a specific action decoder.","The improved tracking accuracy under FCCA (trajectory error roughly halved for moving points) suggests the model learns a more coherent spatio-temporal representation, not just better auxiliary predictions.","Zero-shot robustness on LIBERO-Plus improves most on language and background perturbations and least on camera and robot-state shifts, consistent with the 2D-imagery nature of the tracking supervision."],"fun_headline_variants":["Foresight and tracking boost VLA success to 98.8%","Two training-only objectives push robot policy accuracy to 98.8%","Visual foresight plus motion guidance: key to better VLA actions","Combine future-feature foresight and point tracking for robot control","Training-only foresight and tracking refine robot manipulation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The frozen point-tracker's 2D trajectories are a reliable, action-relevant description of how task objects move; if those pseudo-labels are systematically wrong, the motion guidance baked into the policy is wrong (and the paper itself notes it only models view-dependent motion in image space).","fun_headline_variants_meta":{"raw":{"variants":["Foresight and tracking boost VLA success to 98.8%","Two training-only objectives push robot policy accuracy to 98.8%","Visual foresight plus motion guidance: key to better VLA actions","Combine future-feature foresight and point tracking for robot control","Training-only foresight and tracking refine robot manipulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":2875,"prompt_tokens":745,"completion_tokens":2130,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2041}},"tokens_in":489,"tokens_out":2130,"duration_ms":14735,"temperature":1.0,"reasoning_tokens":2041,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:11:11.349131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FoMoVLA with the point-tracking supervision replaced by scrambled (randomly permuted or inverted) trajectories from the same teacher, keeping all else equal; if LIBERO-Long and RoboCasa success still rise, the motion signal is not what carries the gain.","supporting_citations":[],"review_version":1}