{"id":"1a3cbb8f-9447-4e95-8acc-496657c0c152","arxiv_id":"2607.03693","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Context-conditioned sparse experts in a flow-matching action transformer let a VLA exploit optional depth and specialize by task intent, improving multi-task, long-horizon, and missing-sensor robustness.","lead":"CoRE-VLA makes robot action policies use sparse expert routing so they can use depth when present and still work when depth is missing. It matters because real robots have uneven sensors and long multi-step tasks that shared dense policies handle poorly.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Real-world depth gains may not isolate CoRE routing: physical Orbbec depth is a train–test domain shift relative to DA-V2, so Table 3 improvements could be sensor-domain effects rather than gated experts.","rationale":"The reader correctly flags DA-V2 as the weakest assumption for the unified optional-depth story. I sharpen that to the real-world Table 3 comparison specifically: simulation Table 12 already shows graceful degradation under matched depth on/off (97.6% vs 97.0%), so the central architecture is not internally contradicted. The remaining load-bearing uncertainty is whether physical-depth gains isolate CoRE’s availability-gated experts or mix in train–test depth domain shift. That does not warrant REJECT—the dense DiT ablation and missing-depth real-world rows still support robustness and multi-task gains—but it keeps the verdict CONDITIONAL until a matched pseudo-depth vs physical-depth vs none control is reported (or artifacts allow it). Agreement with the reader is agree on the soft spot; no stronger independent flaw (e.g., routing math, flow-matching objective) overturns the empirical core.","tokens_in":24103,"tokens_out":686,"duration_ms":6234,"concrete_test":"On the same 20-rollout real-world protocol, re-evaluate the identical CoRE-VLA checkpoint under three depth inputs: (A) no depth, (B) DA-V2 pseudo-depth from the live RGB streams (matched to training), (C) raw Orbbec depth. If (B)≈(C) and both beat (A) by a similar margin, the gated-expert claim holds; if (C)≫(B) or (B)≈(A) while (C) improves, the Table 3 physical-depth gains are largely domain-shift and do not cleanly credit modality-specialized experts.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim needs a single checkpoint that (i) exploits auxiliary depth when present and (ii) stays reliable when depth is absent, without sensor-specific fine-tuning. Simulation ablations (Table 12) support this under matched DA-V2 depth on/off. Real-world §4.2 trains only on DA-V2 pseudo-depth from RGB, then deploys the same weights with raw Orbbec Dabai DC1 depth or with depth fully removed. Table 3 shows further gains with physical depth (e.g., Fabric-Folding SR 70%→80%, Clothes-Folding 50%→65%). That comparison confounds two things: enabling modality-specialized experts (z_mod=1, Eq. 5) versus feeding a depth distribution the experts never saw in training. If the lift is mostly domain-shift / calibration / noise-structure mismatch rather than learned geometric specialization, the deployment claim that “modality-specialized experts can exploit geometric cues when available” is only weakly supported outside matched pseudo-depth. The paper’s own design (modality dropout + expert masking) is sound for missing-depth robustness; the load-bearing soft spot is whether physical-depth gains validate the optional-sensor story or only show that a different depth source still helps somewhat.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"CoRE-VLA reformulates VLA action generation as context-conditioned sparse computation inside a flow-matching Action DiT. Task-intent embeddings select a capacity-limited subset of action-side tokens and route them to general or modality-specialized experts, while a modality indicator z_mod (with modality dropout in training and availability-based expert masking) gates auxiliary-modality experts so the same policy can use depth when present and fall back to RGB–language–proprioception when absent. The method is evaluated on LIBERO, RoboCasa GR1 Tabletop, and real dual-arm tasks (Vegetables-Picking, Clothes-Folding, zero-shot Fabric-Folding), with a dense Action DiT ablation and π0.5 as baselines. A detailed LIBERO-Long ablation (Table 12) separates intent routing, naive depth, dropout, and specialized experts; routing visualizations and an efficiency analysis are provided in the appendices.","tokens_in":24490,"tokens_out":1409,"duration_ms":24399,"significance":"The paper addresses a practically important deployment problem—heterogeneous and intermittently missing auxiliary sensors—rather than only scaling a fixed observation interface. Strengths include a clear architecture-and-training formulation (modality-specialized experts + dropout + masking), a carefully factorized LIBERO-Long ablation (Table 12) that isolates over-reliance under naive depth addition, competitive multi-task/long-horizon numbers (LIBERO average 98.7%; Long 97.6% with / 97.0% without depth), real-world with/without depth evaluation with full rollout tables, and qualitative router analyses. If the dual-mode claim holds under cleaner real-world modality matching, the work is a useful template for optional-sensor VLA policies and for intent-adaptive action generators beyond dense shared FFNs.","major_comments":[{"comment":"§4.2 and Table 3: real-world “exploit depth when available” evidence confounds expert gating with a train–test depth domain shift. Training uses only DA-V2 pseudo-depth from RGB; inference with physical Orbbec depth (z_mod=1) improves SR (e.g., Fabric-Folding 70%→80%, Clothes-Folding 50%→65%) relative to depth-off. That lift does not isolate modality-specialized experts learning transferable geometric computation (Eqs. 3–5) from feeding a depth distribution never seen in training. Simulation Table 12 supports matched on/off depth; real-world does not. Please add a matched real-world condition (same checkpoint, DA-V2 depth at inference) and/or qualify claims that physical-depth gains demonstrate learned geometric specialization rather than sensor-domain effects.","section":"§4.2, Table 3"},{"comment":"Table 12 and §3.1: the central dual-mode claim is load-bearing and mostly carried by simulation. Rows 3–4 vs 11–12 cleanly show naive depth over-reliance versus CoRE robustness under matched DA-V2, which is the strongest support for the architecture. The manuscript’s abstract and contribution bullets still lean on real-world physical-depth gains as modality analysis. Align the claim hierarchy with the evidence: state that matched-depth robustness is established on LIBERO-Long, and treat physical-depth deployment as a domain-shift stress test whose positive lift is suggestive but not yet causal for E_mod specialization.","section":"Abstract; Table 12; §3.1"},{"comment":"§4.1 training protocol and contribution 1: “first to explicitly formulate” sensor-heterogeneity as a joint architecture-and-training problem is strong relative to Related Work §2.2–2.3 (auxiliary sensors, ForceVLA-style MoE, ChatVLA MoE). The distinction (availability as an action-generator routing condition with dropout/masking, not only extra inputs) is plausible but should be stated more carefully against concurrent modality-robust and MoE-VLA lines, and limited to the depth instantiation actually studied.","section":"§1 contributions; §2.2–2.3"}],"minor_comments":[{"comment":"Appendix D gives FLOPs bounds for CoRE vs dense blocks but no wall-clock latency or throughput on the same hardware as the dense Action DiT ablation. For a “scalable / sparse computation” framing, even a short measured comparison (action-generator only and full closed-loop) would help readers.","section":"Appendix D"},{"comment":"Real-world evaluation uses 20 rollouts per method/task (Tables 9–11). Reporting binomial CIs or noting variance would make Table 3 score/SR differences easier to interpret, especially on Clothes-Folding where depth-off SR is 50%.","section":"§4.2, Tables 3, 9–11"},{"comment":"Notation: g is defined as AvgPool(C_text) in Eq. (1) and again in the appendix; C vs C(z_mod) and when auxiliary tokens enter cross-attention could be stated once in the main Method with a pointer to Appendix E.","section":"§3.1, Eqs. (1)–(5)"},{"comment":"Figures 3–5 are useful qualitative rollouts; adding failure cases or partial-score modes (already tabulated) in the main text would balance the narrative.","section":"§4.2, Figures 3–5"},{"comment":"Hyperparameters ρ, p_drop, |E_gen|, |E_mod|, CoRE layer indices (Tables 4–5) are free design choices; a short sensitivity note (even one-axis) would strengthen reproducibility claims beyond the fixed settings used.","section":"Appendix A.1"}],"recommendation":"major_revision","confidential_remarks":"Solid systems paper with an unusually careful LIBERO-Long ablation; the main risk for a journal is overselling real-world physical-depth gains as causal evidence for gated experts when training is DA-V2-only. If the authors add a matched DA-V2 real-world inference arm and temper novelty wording, this is close to a clear accept for a robotics methods venue. Scope fit is good for robot learning / VLA deployment; not primarily a theory contribution despite Appendix E."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean robotics systems paper, not a theory one. The useful idea is putting conditional sparse computation inside the flow-matching Action DiT: task-intent Top-K selection of action tokens, general vs modality-specialized experts, and availability masking plus modality dropout so one checkpoint can use depth or drop it without retraining.\n\nWhat is new is the combination and the evaluation, not MoE itself. They already cite ChatVLA/ForceVLA-style routing. The LIBERO-Long ablation (Table 12) is the best part of the paper: it separates dense baseline, intent routing, naive depth, dropout, specialized experts, and even routing on full VL pool vs text intent. That table actually supports the missing-depth story under matched DA-V2 depth (97.6% with / 97.0% without). Real dual-arm results beat a dense DiT and fine-tuned π0.5, including zero-shot fabric folding. Routing heatmaps and efficiency notes are honest extras, not decoration.\n\nSoft spots, in proportion. Training depth is DA-V2 pseudo-depth; real inference either removes depth or plugs in Orbbec raw depth. Table 3’s further gains with physical depth therefore mix “experts on” with a train–test depth domain shift. The stress-test is fair: that does not kill the missing-depth claim (RGB-only still beats dense), but it weakens the clean story that modality experts “exploit geometric cues” on physical sensors without fine-tuning. Multi-sensor generality is mostly aspirational—experiments are depth-only. “First formulation” language is a bit strong given prior VLA MoE. Free knobs (ρ, expert counts, CoRE layers, dropout) are many; no code/data release yet; main tables could use variance like RoboCasa’s.\n\nMath is standard flow matching plus usual MoE/selection regularizers—fine for this venue. Citations look appropriate. Who cares: people building multi-platform VLAs who actually hit sensor heterogeneity. I would bring it to reading group and send it to referees. Ask for tempered claims, clearer physical-depth controls if possible, and artifacts. Worth engaging.","headline":"Solid systems paper: intent- and sensor-gated sparse action experts with real missing-depth robustness; physical-depth gains are a bit confounded but not fatal.","tokens_in":25129,"tokens_out":550,"would_cite":true,"duration_ms":5906,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Robot action generation can be sparse and context-conditioned so one policy uses optional sensors when present and still acts when they are missing.","keywords":["Vision-Language-Action Models","Robotic Manipulation","Conditional Routing","Sparse Computation","Multimodal Policy Learning","Mixture of Experts","Depth Sensing","Flow Matching"],"falsifier":"Train the full CoRE model with estimated depth only, then compare the same real dual-arm tasks with depth removed versus with a physical depth camera and versus a dense no-depth action generator; a large drop without depth, or no gain (or harm) from physical depth, would falsify graceful optional-modality routing.","tokens_in":24980,"feed_emoji":"🤖","tokens_out":895,"duration_ms":17138,"temperature":0.7,"pith_summary":"Robots in the wild do not share one sensor suite: some have depth or other auxiliaries, others lack them by design, and sensors that exist can fail mid-run. Standard vision-language-action policies still couple action generation to a fixed observation set through shared dense computation, so they over-rely on auxiliaries or cannot specialize across diverse and long-horizon tasks. CoRE-VLA instead treats action generation as sparse computation conditioned on two kinds of context: whether an auxiliary modality is available, and what the language instruction currently requires. Sensor availability gates modality-specialized experts; task intent selects which action-side representations receive expert computation. On multi-task and long-horizon simulation suites and on real dual-arm picking and folding, the same checkpoint gains from depth when it is present and keeps high success when depth is removed, without a second training run.","feed_headline":"One robot policy uses depth when present, survives without it","feed_subtitle":"Sparse expert routing by task intent and sensor availability beats dense action generators on long-horizon work.","key_machinery":"Conditional Routing of Experts (CoRE): inside the action diffusion transformer, a task-intent embedding sparsely selects action-side tokens and routes them to general or modality-specialized experts, while a modality indicator masks the specialized experts whenever the auxiliary sensor is off.","core_discovery":"Formulating VLA action generation as context-conditioned sparse computation—task-intent selection and routing of action-side tokens plus availability-gated modality-specialized experts with modality dropout—yields a unified policy that exploits auxiliary depth when available, remains reliable when depth is absent without retraining, and improves multi-task and long-horizon success over a dense action-generator ablation and a strong pretrained VLA baseline.","pith_inferences":["If expert mixtures and token-selection patterns track subgoals over a rollout, those routing traces could become a free diagnostic signal for which stage of a long-horizon failure the policy is in.","One shared checkpoint with availability masks could serve heterogeneous multi-robot fleets that differ only in which cameras they carry.","Keeping activated compute proportional to a fixed selection ratio while growing expert count suggests the design scales more cleanly as action horizons and embodiment action dimensions grow.","A small gap between estimated-depth training and physical-depth inference would imply the specialized experts capture geometry-like structure rather than sensor-specific noise—worth testing with other depth estimators."],"forward_implications":["An RGB-pretrained VLA can be extended with optional sensors by continued training without redesigning the full action backbone.","Missing-sensor deployment no longer needs a separate policy or sensor-specific fine-tuning pass.","Long-horizon multi-task learning can reduce interference by assigning different subgoals to different expert paths.","A physical depth camera can be enabled at inference even when training used only estimated depth.","The same gating pattern is intended to host other optional physical modalities such as touch or force."],"fun_headline_variants":["CoRE-VLA routes experts by task and sensor for robust robot action","Sparse experts let one VLA use depth or drop it without retraining","Task-intent and sensor-gated experts beat dense VLA on long horizons","Availability-gated modality experts keep policies solid when sensors fail","Context-conditioned sparse routing unifies multi-task VLA under missing depth"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that depth estimated from RGB during training is close enough to real depth cameras that the modality experts learn transferable geometric skills usable with physical depth or with depth fully removed.","fun_headline_variants_meta":{"raw":{"variants":["CoRE-VLA routes experts by task and sensor for robust robot action","Sparse experts let one VLA use depth or drop it without retraining","Task-intent and sensor-gated experts beat dense VLA on long horizons","Availability-gated modality experts keep policies solid when sensors fail","Context-conditioned sparse routing unifies multi-task VLA under missing depth"]},"model":"grok-4.5","effort":"low","cost_usd":0.00394,"raw_usage":{"total_tokens":1294,"prompt_tokens":860,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":39400000,"prompt_tokens_details":{"text_tokens":860,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":351,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":860,"tokens_out":83,"duration_ms":3136,"temperature":1.0,"reasoning_tokens":351,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:37:27.784114+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the full CoRE model with estimated depth only, then compare the same real dual-arm tasks with depth removed versus with a physical depth camera and versus a dense no-depth action generator; a large drop without depth, or no gain (or harm) from physical depth, would falsify graceful optional-modality routing.","supporting_citations":[],"review_version":1}