{"id":"7ab9c364-0424-4778-9a28-2fff3e80be7b","arxiv_id":"2602.02459","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A latency-aware vision-language-action model that conditions control on deliberately delayed semantic features and latency metadata outperforms prior VLA navigation models under multi-second reasoning delays.","lead":"TIC-VLA is a robot navigation system that lets slow AI language reasoning run in the background while a fast controller drives the robot, explicitly telling the controller how old the reasoning is. It reports better success in simulated and real indoor/outdoor tests than prior vision-language-action models, especially on low-power hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IL stage trains action policy on KV caches computed from teacher-forced ground-truth reasoning, so the reported interface ablations may not reflect deployment; the final RL stage's dependence on the interface is unablated.","rationale":"I disagree with the reader's weakest_assumption. The synthetic delay distribution is a reasonable first-order model and is directly supported by real-world deployment: the trained model achieves 75–85% success on edge hardware with actual latencies of 3.4–4.8 s, while training delays were U(0,10) and ~5 s. Covariate shift in the observed latency scalar is unlikely to be fatal because the policy conditions on Δt explicitly. The more serious concern is an internal training/deployment mismatch in the semantic inputs. The supplementary material explicitly states that IL uses ground-truth reasoning annotations, and Algorithm 1 feeds R_{t−Δ} into the VLM when extracting KV caches. At test time the VLM must generate its own reasoning, so the action policy is trained on a different distribution of semantic features. The paper's core evidence for the delayed semantic-control interface (Table 2, Table 9) comes from this IL-only setup, so those ablations conflate the interface's benefit with access to ground-truth future-informed reasoning. The RL stage uses genuine VLM outputs but only tunes the last layers and lacks post-RL interface ablations. This is a concrete, checkable methodological gap rather than a matter of consensus. It does not invalidate the overall system—the real-world results show the full pipeline works—but it means the central causal claim is not yet established. Thus the conditional verdict is appropriate; no change needed.","tokens_in":26480,"tokens_out":12377,"duration_ms":119013,"concrete_test":"Modify Algorithm 1 to obtain the KV cache from the frozen VLM's own free generation on delayed observations (no ground-truth reasoning input). Replace line 5 with S_{t−Δ} ← f_LM(V_{t−Δ}^{vlm}, I), letting the VLM autoregressively generate its reasoning and waypoints before caching the KV. Train the action expert from scratch under the same latency sampling and evaluate on the 85 DynaNav episodes. Compare SR/CR/NE/SPL against the reported no-RL numbers (47.06/34.12/10.85/42.41). If SR drops to near the waypoint-interface level (~22–30), the IL ablation evidence is inflated; if it remains ~47, the teacher-forcing is benign. As a secondary check, evaluate the final RL-finetuned TIC-VLA on DynaNav with the KV-cache input removed (replaced with zeros/random vectors) and with latency metadata removed, to test whether the interface still matters after RL.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 1 (line 5), the KV cache used to train the action expert is obtained as S_{t−Δ} ← f_LM(V_{t−Δ}^{vlm}, I, R_{t−Δ}), where R_{t−Δ} is the ground-truth GPT-5 reasoning trace generated offline from future observations (Supplementary B: 'the policy is conditioned on delayed ground-truth reasoning annotations and guidance waypoints'). At deployment, no R is provided; the VLM must generate its own reasoning from current/past observations. Thus the IL objective conditions the policy on semantic features that encode future information and gold-quality reasoning, violating the stated 'latency-consistent training' principle at the semantic level. The latency distribution is matched, but the semantic content is not. Since the component analyses in Table 2 (KV Cache vs Waypoint, with/without latency) and Table 9 (ego-motion offset) are reported for the IL-only variant (SR 47.06 matching TIC-VLA (no RL) in Table 1), these results likely overstate the benefit of the delayed semantic interface and latency metadata. The RL stage (Algorithm 2) does use free VLM generations and is therefore deployment-like, but it fine-tunes only the final cross-attention layer and MLP of the action expert, and the paper reports no post-RL ablations of the interface. Consequently, the evidence that the delayed semantic-control interface is the load-bearing component is incomplete.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TIC-VLA, a vision-language-action framework for robot navigation that explicitly models the latency of VLM reasoning. The core idea is a delayed semantic-control interface: the action policy conditions on the KV-cache features of a previous VLM reasoning update, along with explicit latency and ego-motion offsets, while receiving current visual observations at 10 Hz. Training proceeds in three stages: VLM SFT on GPT-5-annotated reasoning traces, imitation learning with injected uniform [0,10] s delays, and PPO fine-tuning under an average 5 s injected delay. The authors introduce DynaNav, a new Isaac Sim-based benchmark with 85 episodes across four scenes and dynamic pedestrians. Simulation results (Table 1) show TIC-VLA with higher SR (55.29) and lower CR (28.24) than Uni-NaVid, NaVILA, and DualVLN; real-world tests on a Unitree Go2 report 85% average success across four tasks with 5 trials each. Ablations examine the semantic interface, latency metadata, ego-motion offsets, action policy architecture, and VLM backbone.","tokens_in":26873,"tokens_out":3644,"duration_ms":34970,"significance":"If the central claim holds, TIC-VLA addresses a real and underexplored deployment bottleneck: multi-second VLM reasoning latency in closed-loop control. The idea of treating latency as a first-class modeling variable, rather than an engineering nuisance, is timely, and the explicit pseudo-code and DynaNav benchmark are useful contributions. The paper also reports real-robot results on two edge platforms, which is commendable. However, the evidence for the load-bearing architectural claim is currently incomplete: the IL-stage ablations use ground-truth delayed reasoning traces that encode future information, and the post-RL stage does not re-ablate the interface. With no error bars or significance tests anywhere, and only 5 trials per real-world task, the quantitative claims are plausible but not yet rigorously supported.","major_comments":[{"comment":"The IL stage conditions the action policy on ground-truth delayed reasoning traces R_{t-Δ}, which are generated offline from future observations (Supp. B: 'the policy is conditioned on delayed ground-truth reasoning annotations and guidance waypoints'). At deployment, R is generated by the VLM from current/past observations only. The latency distribution is matched, but the semantic content is not. Because the interface ablations in Table 2, Table 9, Table 10, and Table 11 are reported for the IL-only variant (SR 47.06 matching TIC-VLA (no RL) in Table 1), these results likely overstate the benefit of the delayed KV-cache interface and latency metadata. The RL stage (Algorithm 2) uses free VLM generations, but only fine-tunes the final cross-attention layer and MLP of the action expert; no post-RL ablations of the interface are reported. Thus the central claim that the delayed semantic-c","section":"§3.3, Algorithm 1, Supp. B"},{"comment":"No error bars, confidence intervals, or significance tests are reported anywhere. Real-world results are based on only 5 trials per task; for example, in Supp. Table 8 the Outdoor Campus comparison (TIC-VLA 80% vs DualVLN 60% vs NaVILA 60%) corresponds to 4/5 vs 3/5 vs 3/5 successes, which is well within chance-level variation. Figure 5 is presented without numerical values or error bars. At minimum, report per-episode results and confidence intervals for real-world tests, and mean±std over multiple seeds for the simulation benchmark.","section":"Table 1, Table 3, Supp. Table 8, Figure 5"},{"comment":"The baselines are described as 'fine-tuned on the same datasets to ensure a fair and controlled comparison', but Supp. D only states that 'we follow their respective open-source implementations and training'. No details are given on the amount of fine-tuning data, number of epochs, learning rates, or how the discrete-action or mid-level-action baselines are adapted to the 10 Hz continuous control setting. This is essential for assessing the fairness of Table 1 and Table 3. Please specify the exact fine-tuning protocols, including data mixing, training steps, and action-space adaptation.","section":"§4.1, Supp. D"},{"comment":"The DynaNav benchmark is authored by the same group and built in the same Isaac Sim environment used for IL/RL training, and the latency-robustness result is tested under the same type of injected delay used during training (uniform [0,10] s in IL, ~5 s in RL). This is not circular in a logical sense, but it raises a correctness-risk concern: the claimed robustness may be tuned to the authors' specific delay distribution. Please provide at least one evaluation under a held-out delay distribution (e.g., a truncated Gaussian, or delays measured from actual Jetson/GPU inference profiles) or on an independent benchmark to show that the method transfers.","section":"§3.4, §4.2"}],"minor_comments":[{"comment":"Typo: 'historical trajecoty' should be 'historical trajectory'.","section":"§3.1"},{"comment":"Typos: 'Drouput rate-KV cache' and 'Drouput rate-state' should be 'Dropout rate-KV cache' and 'Dropout rate-state'.","section":"Table 5"},{"comment":"Inconsistent naming: 'Dual-VLN' in Table 3/Table 8 vs 'DualVLN' in the text; 'NA VILA' in Supp. D should be 'NaVILA'.","section":"Supp. D, Table 8"},{"comment":"In the middle example, the text says 'passing the water fountain' and then 'stop by the fountain', but the instruction is to stop at the water dispenser; the duplicate phrase 'water water fountain' also appears. Please clarify the target and fix the wording.","section":"Figure 11"},{"comment":"The paper states 'we fine-tune only the action policy using RL while keeping the vision encoder and language model frozen', but a later sentence says 'We finetune only the final cross-attention Transformer layer and the MLP decoder of the action expert'. Please reconcile these descriptions.","section":"Supp. D"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what this paper actually adds: it treats VLM inference latency as a variable the control policy can consume. The delayed semantic-control interface — caching KV features and passing the delay plus ego-motion offset to the action expert — is a genuine step past previous dual-system VLAs, which decouple but assume fresh semantics. The latency-consistent training (injecting delays in IL and RL) is the right instinct, and the simulation results are directionally strong: TIC-VLA roughly doubles SR against the strongest baselines. The real-robot numbers, even with five runs per task, support the claim that asynchronous control helps on edge hardware. That part deserves credit.\n\nThe soft spots are not fatal but they matter. The biggest is a training/deployment mismatch in the IL stage. Algorithm 1 conditions on S_{t-Δ} computed from ground-truth reasoning traces generated from future observations (Supp B). That is not the same semantic content the policy will see at deployment, where the VLM must generate reasoning from current/past observations only. So the component ablations in Table 2 and Table 9, which are IL-only, likely overstate the benefit of the delayed KV interface and latency metadata. The RL stage uses free generations and is more honest, but it fine-tunes only the last cross-attention layer, and no post-RL ablation of the interface is reported. The central load-bearing claim is therefore not fully nailed down.\n\nSecond, the evaluation lacks error bars and significance tests; 5 real-world trials give an 80% vs 50% difference that is suggestive, not robust. Third, the text says reasoning 'improves all navigation metrics' but the CR in Table 4 worsens (25.88 to 28.24). Fourth, the main text says action LR is 2e-4, Table 6 says 2e-5. Minor, but sloppy. Fifth, DynaNav is built by the same group and the same simulator is used for IL/RL training, so the benchmark is useful but not independent.\n\nWho is this for? Groups building real VLA navigation on edge robots; it is a practical subfield-level improvement, not a world-changer. With the evidence gap fixed — post-RL ablations, error bars, a code/data release, and consistency fixes — it would be a solid conference paper. As is, it deserves peer review, not desk rejection, but the review should push on the IL-to-deployment gap.","headline":"A useful latency-aware VLA navigation system with a real design idea, but the core evidence for the interface is weakened by an IL-stage training/deployment mismatch and thin statistics.","tokens_in":27338,"tokens_out":3002,"would_cite":false,"duration_ms":29156,"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 a robot control policy can safely act on vision-language reasoning that is seconds out of date, if it is told exactly how stale that reasoning is and by how much the robot has moved since.","keywords":["vision-language-action models","robot navigation","inference latency","delayed semantic-control interface","asynchronous reasoning","imitation learning","reinforcement learning","dynamic environments"],"falsifier":"Measure success of the trained TIC-VLA policy under a latency distribution whose support extends well beyond 10 seconds (e.g., 15-25 s) or with a bi-modal delay profile; if success rate does not degrade relative to the 0-10 s range, the latency-conditioning hypothesis is not doing the claimed work. Similarly, if a policy trained with no injected delay matched the latency-trained policy under multi-second delays, the paper's assertion that latency-consistent training is required would be falsified.","tokens_in":26402,"feed_emoji":"🤖","tokens_out":5504,"duration_ms":49609,"temperature":0.7,"pith_summary":"This paper attempts to prove that inference latency in language-guided navigation is not an engineering nuisance to be hidden or waited out, but a quantity the controller should be handed explicitly. It proposes TIC-VLA, in which a slow vision-language model periodically emits semantic features from a past frame, and a fast action policy consumes those stale features together with the elapsed delay and the robot's motion during that delay. The paper's central argument is that training the policy under delays matching deployment makes it robust to multi-second thinking time, while synchronous training collapses. Evidence is offered in a physics-based simulator and on a real quadruped, where the approach reports higher success rates than VLA baselines that assume temporally fresh reasoning.","feed_headline":"Robot navigates safely even when its AI 'thinking' is seconds late","feed_subtitle":"Exposing how stale semantic predictions are and training with injected delays pushes real-robot success to 85%.","key_machinery":"The central mechanism is the delayed semantic-control interface. The VLM's last-layer key-value cache — the internal attended features of its final forward pass — is saved and re-served to the action policy at every control step until a new inference completes. Alongside it, the policy receives the effective delay Δt (inference time plus elapsed time since completion) and the ego-motion displacement (Δx,Δy,Δθ) accumulated since the reasoning was produced, letting the transformer action expert locate the stale semantics in the current body frame. The second half of the mechanism is latency-consistent training: the same delayed-features-plus-metadata conditioning is synthesized during imitatio","core_discovery":"TIC-VLA decouples slow semantic reasoning from fast reactive control through a delayed semantic-control interface: the action policy receives the last completed VLM KV-cache features, the effective reasoning latency Δt, and the ego-motion offset accumulated during that delay, in addition to current images and robot state. The VLM is anchored to a past observation, the policy is trained with delays sampled uniformly from 0 to 10 seconds during imitation learning and roughly 5-second delays during reinforcement learning, and at test time the policy never blocks on inference. On the DynaNav benchmark the full model reports 55.29% success against 22-31% for language-conditioned baselines, and on","pith_inferences":["The same timestamp-and-motion-offset interface could generalize to other closed-loop systems with periodic slow reasoning, such as manipulation planners or assistive driving, where a dense latent state from the past is available.","Rather than a fixed vision-language update schedule, a learned gating policy could decide when the expected information gain of a new inference outweighs the staleness it introduces.","If the deployment latency distribution differs from the uniform 0-10 s training range (e.g., heavy-tailed or correlated delays), the policy may need online adaptation or adversarial delay sampling to keep its robustness."],"forward_implications":["Language-guided navigation can run on edge hardware with a small model and a slow vision-language model, without pausing the robot while the model thinks.","Policies trained under synchronous, fresh reasoning will degrade under realistic inference latency; latency injection should be part of standard VLA training.","Dense semantic features (KV cache) are a more effective interface for delayed guidance than sparse high-level waypoints.","Reinforcement fine-tuning under injected delays further stabilizes success as inference latency grows and also lowers collision rate.","Explicit ego-motion compensation allows a stale semantic state to be re-interpreted in the robot's current coordinate frame."],"fun_headline_variants":["Robot navigates despite seconds of AI thinking lag","TIC-VLA: robot control that outpaced its own brain","Navigation robot tolerates delayed semantic reasoning","Latency-aware robot handles slow AI without stalling","Robot moves safely even when its AI thinks late"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that delays sampled uniformly from 0 to 10 seconds in imitation learning and roughly 5 seconds in reinforcement learning faithfully represent the real inference-latency distribution and ego-motion error of edge hardware; if actual latency profiles, scheduling jitter, or odometry noise are very different, the claimed robustness may not carry to deployment.","fun_headline_variants_meta":{"raw":{"variants":["Robot navigates despite seconds of AI thinking lag","TIC-VLA: robot control that outpaced its own brain","Navigation robot tolerates delayed semantic reasoning","Latency-aware robot handles slow AI without stalling","Robot moves safely even when its AI thinks late"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1100,"prompt_tokens":737,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":481,"tokens_out":363,"duration_ms":4188,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:20:15.414055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure success of the trained TIC-VLA policy under a latency distribution whose support extends well beyond 10 seconds (e.g., 15-25 s) or with a bi-modal delay profile; if success rate does not degrade relative to the 0-10 s range, the latency-conditioning hypothesis is not doing the claimed work. Similarly, if a policy trained with no injected delay matched the latency-trained policy under multi-second delays, the paper's assertion that latency-consistent training is required would be falsified.","supporting_citations":[],"review_version":1}