{"id":"fa115dc9-b81c-4fea-8c16-600f0df8e5b4","arxiv_id":"2504.18057","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A cloud-vehicle system that triggers a large vision model only when local detection confidence is low, and uses forward simulation to decide whether cloud guidance will actually improve the MPC trajectory, improves navigation time and success rate in CARLA.","lead":"This paper proposes a system that lets a self-driving car ask a cloud-based large vision model for help only when its local sensors are uncertain, using the cloud's advice to guide a local controller. In CARLA simulations, the system finishes routes faster than local-only or fixed-interval cloud plans while using fewer cloud queries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CFS service decisions hinge on an uncalibrated uniform distribution for post-cloud confidence; without calibration or sensitivity analysis, the claimed timing advantage is unsupported.","rationale":"The reader's weakest assumption correctly identifies the CFS confidence distribution as the most load-bearing unvalidated step. The central empirical claim is that OCP outperforms LOS and PCS in navigation time, success rate, and query efficiency; the query-efficiency and timing advantages are produced specifically by CFS's accept/reject decisions, which are computed from Eq. (12) under an assumed uniform distribution centered at Cthreshold. No calibration data, hyperparameter sensitivity, or ablation is provided to show that this assumption is harmless. The paper's CARLA experiments provide some independent support for the overall system, but they are few in number, lack error bars in key comparisons, and do not isolate the CFS decision rule. The false unimodality remark and the contradictory success-rate sentence ('reduces the success rate') are real defects but not the central correctness risk. A direct recalibration check would settle whether the CFS assumption is benign or whether the claimed timing advantage depends on an ungrounded prior.","tokens_in":12217,"tokens_out":3943,"duration_ms":45024,"concrete_test":"Collect the cloud LVM confidence scores c^diamond_{m,t} for all objects that triggered a query in Experiments 2-4 (or on the test set's unknown classes), estimate the empirical distribution, then re-run CFS with this empirical P in place of U(Cthreshold +/- Delta_c). If the service decision sequence beta_t changes in any reported scenario, or if final FTime or TLenh shift by more than a few percent, the current uncalibrated distribution is load-bearing and the timing claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The CFS block (Sec. V-B) is the mechanism that accepts or rejects every cloud service, and it does so using E[Delta_J] computed from an assumed distribution P = U(Cthreshold - Delta_c, Cthreshold + Delta_c). This is an assumption, not an estimate: no data from the cloud LVM on unknown objects is used to set its center, width, or shape, and no sensitivity analysis is reported. If true post-redetection confidences are systematically above the threshold (the natural case for a correct LVM redetection) or systematically below it (frequent LVM failures), the expected trajectory improvement in Eq. (12) is biased in a definite direction. That bias directly changes the accept/reject decisions that constitute the CTO contribution, and therefore changes the claimed advantage over fixed-period PCS. The experiments do not distinguish between 'CFS timing is correct' and 'CFS timing happens to be correct in these few hand-built scenarios.' The unimodality claim after Eq. (10) is also false as stated (the sum of an increasing and a decreasing step function need not be unimodal), but it is a supporting remark, not the load-bearing step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Opportunistic Collaborative Planning (OCP), a cloud-vehicle collaborative navigation system for open scenarios with unknown objects. OCP couples a local detector and model predictive controller with a cloud large vision model (LVM) pipeline: the cloud uses SAM-based perception to recalibrate detections, infer object-lane topology, and generate behavior decisions and waypoints that reconfigure the local MPC (LVM-MPC). A collaboration timing optimization (CTO) block decides when to query the cloud via object detection confidence thresholding (ODCT) and when the cloud should accept the query via cloud forward simulation (CFS), which estimates expected trajectory improvement after redetection. The system is implemented in CARLA/ROS and compared with local-only (LOS) and fixed-period collaboration (PCS) baselines in static and dynamic obstacle scenarios; the paper reports shorter finish times, higher success rates, and fewer cloud invocations for OCP.","tokens_in":12384,"tokens_out":5850,"duration_ms":64382,"significance":"If the empirical claims hold, the paper offers a practical mechanism for deciding when to offload perception and planning to a cloud LVM, which is a real bottleneck for resource-constrained autonomous systems. The main contribution is the CTO gating logic, since the reported gains over fixed-period PCS come from avoiding unnecessary cloud calls while still obtaining timely assistance. The CARLA implementation covers static multi-obstacle scenarios, open-set detection with strict train/validation/test class splits, and a dynamic obstacle scenario, which is a reasonable breadth for a robotics systems paper. The paper does not provide open-source code or machine-checked proofs, so its contribution is empirical and would be strengthened by releasing the implementation; nonetheless, the system-level integration and the ODCT validation protocol are explicit strengths.","major_comments":[{"comment":"The CFS service decision, which is the component that distinguishes OCP from fixed-period PCS, is computed under the assumed post-cloud confidence distribution P = U(Cthreshold - Delta_c, Cthreshold + Delta_c) with mean Cthreshold and fixed width Delta_c = 0.1. This distribution is not calibrated against any measured post-redetection LVM confidences, and no sensitivity analysis is reported. If actual LVM confidences on unknown objects are systematically above Cthreshold (successful redetection) or below it (failed redetection), the expected trajectory improvement E[Delta_J] in Eq. (12) is biased in a definite direction, and the accept/reject decisions that produce the reported reduction in cloud services are not grounded. The authors should either estimate P from data or demonstrate that the OCP timing advantage is robust across plausible alternative distributions.","section":"V-B, Eq. (12)"},{"comment":"The success-rate reporting is internally contradictory: the text states 'Compared to the second-bast scheme, the proposed OCP reduces the success rate by over 6%,' while Fig. 9 and the preceding sentences describe OCP as having the highest success rate, close to 100%. Either the authors mean that the failure rate is reduced by over 6%, or the success-rate advantage claimed in the Abstract is misstated. As printed, the headline success-rate result is ambiguous and must be corrected.","section":"VI-C, Fig. 9 and text below Table III"},{"comment":"The quantitative performance claims are presented only as point estimates: 10 runs for Table II, 30 runs for Table III, and an unreported number of runs for Table IV, with no standard deviations, confidence intervals, or significance tests. Since CARLA is stochastic and some baselines fail partially, the reported differences between OCP and PCS (e.g., 22.35 s vs. 24.09 s in Table II and 17.28 s vs. 23.40 s in Table IV) cannot be separated from run-to-run variability. Please report dispersion and significance, or explicitly qualify the results as illustrative single-scenario demonstrations.","section":"VI-B through VI-D, Tables II-IV and Fig. 9"}],"minor_comments":[{"comment":"The statement that G(Cthreshold) is 'guaranteed to be unimodal' is not correct as stated: the sum of a monotonically increasing step function and a monotonically decreasing step function need not be unimodal. Please replace this claim with a description of the grid-search optimization actually used, or prove unimodality under additional conditions.","section":"V-A, Eq. (10)"},{"comment":"The number N of Monte Carlo samples used in Eq. (12) is never specified, and the indexing in the second sum of Eq. (13) appears inconsistent (the second sum runs from i = L rather than i = t + L). Please clarify both points, as they affect the reproducibility of the CFS implementation.","section":"V-B, Eqs. (12)-(13)"},{"comment":"There are typos in this section: 'Cthrehold' should be 'Cthreshold,' and 'the second-bast scheme' in Section VI-C should be 'the second-best scheme.' Also, the phrase 'fine-tuned on the validation set to optimize Cthreshold' should be rephrased to make clear that only the threshold is selected on the validation set, not that the detector is trained on it.","section":"VI-A"},{"comment":"The text contains 'the the Carla Town03 map' and states that 'LOS always fails' without numeric support; please fix the typo and provide at least the success-rate count for LOS in the dynamic scenario.","section":"VI-D"},{"comment":"The Introduction indicates that OCP is compared with the collaborative planner [9], but the experiments only compare LOS, PCS, and OCP; the statement that [9] 'has a similar performance as LOS' is not backed by any experiment in the paper. Either add the comparison or remove the claim.","section":"II and VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript cites several works from the same group (RDA, Neupan, and the collaborative planner [9]) and relies on them as baselines or components; this is understandable given the direct lineage, but the editors may wish to ensure that the similarity claim about [9] is either verified or softened. My recommendation is based on the manuscript's own evidence, particularly the uncalibrated CFS distribution and the missing variability reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible systems paper with a genuinely useful idea—deciding when to invoke a cloud vision model based on predicted trajectory improvement. The integration of SAM-based perception, confidence-aware MPC, and a forward-simulation service policy is new, and the CARLA results mostly support the claim that opportunistic cloud guidance beats fixed-schedule collaboration. But the service-decision module, CFS, assumes a uniform distribution for post-cloud confidences centered at the threshold, with no calibration or sensitivity analysis. That assumption is the mechanism that accepts or rejects every cloud query, so the reported timing advantage is not yet grounded. The unimodality claim in Eq. (10) is false as stated (sum of a step up and step down need not be unimodal), though it is a remark, not load-bearing. Also, the sentence in Sec. VI-C saying OCP 'reduces the success rate by over 6%' contradicts the preceding near-100% figure; must be a typo for failure rate. The experiments report averages over 10-30 runs but no variance or significance tests, so we can't tell if the finish-time differences are real. They also exclude failed runs from the trajectory metrics, which is common but worth stating.\n\nWhat's genuinely good: the architecture is clean, the MPC formulation with confidence-dependent inflation is sound, and the experimental setup (unknown objects in CARLA) is appropriate. The ODCT threshold validation on a separate validation set is standard model selection, not circular. Self-citation is heavy (RDA, prior collaborative planning) but the cited pieces are real building blocks, not invented.\n\nWho should read it: people building cloud-robot systems, and MPC researchers interested in perception-triggered receding-horizon control. It deserves a proper peer review, but it needs revision before accept: release code, add error bars, calibrate or sensitivity-test the CFS distribution, and fix the success-rate wording. I'd send it out.","headline":"Useful integration of cloud LVM guidance with local MPC, but the timing policy's service decisions rest on an uncalibrated confidence distribution and the experiments need error bars.","tokens_in":12980,"tokens_out":2273,"would_cite":true,"duration_ms":23399,"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 claims that an opportunistic planner that decides when to query a cloud large vision model and when to accept its service reduces navigation time by over 26 percent and keeps success rate near 100 percent in unknown-object…","keywords":["collaborative planning","large vision model","model predictive control","cloud robotics","unknown object detection","confidence thresholding","forward simulation","autonomous navigation"],"falsifier":"Collect a test set of images with unknown objects, run the cloud detector on objects the local model flags as low-confidence, and compare the resulting confidence distribution to the assumed uniform band around $C_{\\text{threshold}}$; also compare CFS's accept/reject decisions against whether cloud guidance actually shortens the MPC trajectory in the simulator.","tokens_in":11965,"feed_emoji":"🚗","tokens_out":6233,"duration_ms":59422,"temperature":0.7,"pith_summary":"This paper sets out to solve a coordination problem: autonomous vehicles need large vision models to recognize rare objects, but calling the cloud too often wastes resources and can even hurt. It claims that the right policy is to couple a local model predictive controller with cloud-based large vision model guidance, and to decide collaboration timing by optimizing a confidence threshold and simulating the expected trajectory gain before serving a query. If the claim holds, vehicles can handle unknown objects nearly as well as a fully cloud-driven system while querying the cloud only when it pays off. The evidence is presented in simulated urban driving, where the proposed scheme finishes trips over 26 percent faster than fixed-interval collaboration and maintains close to 100 percent success.","feed_headline":"Cloud queries timed by simulation cut navigation time 26%","feed_subtitle":"A hybrid planner calls a large vision model only when low-confidence objects appear and a forward simulation says it will pay off.","key_machinery":"The central object is the collaboration state $\\beta_t \\in \\{0,1\\}$, a binary switch that reconfigures the local MPC's collision-avoidance inflation distance, reference waypoints, and control style between local detections ($\\beta=0$) and cloud detections ($\\beta=1$). The timing machinery that sets $\\beta_t$ is CTO: ODCT picks a confidence threshold $C_{\\text{threshold}}$ by maximizing the sum of recall on unknown and known detections, and CFS runs multiple forward simulations of the MPC under sampled post-cloud confidence scores from a uniform distribution centered at the threshold, computing expected trajectory improvement while accounting for latency; the cloud serves the query only if that improvement is positive.","core_discovery":"The paper's central claim is that an autonomous vehicle can navigate open scenes containing rare, previously unseen objects by pairing a small local model predictive controller with a large cloud vision model, and by deciding the collaboration timing explicitly. In LVM-MPC, the cloud re-perceives the scene with a zero-shot segmentation model, extracts lane and object relations, and generates behavior decisions, waypoints, and control-style reconfigurations that guide a local MPC; the MPC remains feasible under vehicle dynamics and collision-avoidance constraints. In CTO, the vehicle first sets a confidence threshold to decide when to query the cloud, and the cloud then runs forward simulations of the MPC under sampled post-redetection confidences to estimate expected trajectory improvement, accounting for communication and computation latency, and serves only when that improvement is positive. In simulated urban driving, the integrated scheme finishes trips over 26 percent faster than fixed-interval collaboration in a dynamic-obstacle scenario, keeps success rate close to 100 percent, and uses about half as many cloud services.","pith_inferences":["Beyond the paper, the uniform-confidence assumption in CFS could be replaced with a learned predictive model of post-redetection confidence, which might make accept/reject decisions robust without additional experiments.","The same query-service optimization could extend to multi-vehicle fleets sharing one cloud LVM, where the cloud must also schedule queries across vehicles.","The forward-simulation cost of CFS, which solves the MPC many times per query, could become a bottleneck at scale; an amortized or learned approximation of expected trajectory improvement is a natural next step.","The threshold optimization objective could be extended to include query cost and latency explicitly, rather than only detection recalls."],"forward_implications":["A vehicle can navigate unknown-object scenes with cloud assistance roughly as well as a continuously cloud-guided system while cutting the number of large-model services in half.","The optimized confidence threshold generalizes from validation to test sets even when the unknown classes are disjoint, so unseen-object detection is triggered without retraining.","CFS rejects queries when redetection would not change the planned trajectory, so cloud resources are spent only when they move the vehicle.","In dynamic scenes, a single well-timed cloud service can enable an overtake that fixed-interval collaboration misses, cutting finish time by more than 26 percent.","Success rate stays near 100 percent in the tested scenarios, whereas local-only and fixed-interval planners degrade as unknown objects increase."],"supporting_citations":[{"why":"Supplies the RDA model predictive controller and the vehicle dynamics used in the LVM-guided MPC problem.","marker":"[3]"},{"why":"Supplies the zero-shot segmentation model that generates the masks for cloud perception and lane detection.","marker":"[17]"},{"why":"Supplies the local detection architecture that produces the confidence scores used by ODCT.","marker":"[18]"},{"why":"Supplies the open urban driving simulator in which all experiments are run.","marker":"[16]"},{"why":"Defines the periodical cloud collaboration baseline (PCS) that OCP is compared against.","marker":"[12]"},{"why":"Defines the collaborative motion planning baseline that behaves like local-only planning without LVM guidance.","marker":"[9]"},{"why":"Provides the uncertainty-aware inflation distance formulation reused in the collision-avoidance constraint.","marker":"[10]"},{"why":"Provides forward simulation for interactive planning, the basis of CFS's trajectory improvement estimate.","marker":"[20]"},{"why":"Provides forward simulation for exploration planning, cited as the basis for CFS's simulation approach.","marker":"[21]"}],"fun_headline_variants":["LVM-guided MPC saves 26% trip time in open driving","Simulation-timed cloud calls speed up autonomous driving","Collaborative planning: local MPC + cloud LVM, 26% faster","Opportunistic cloud queries cut navigation time by 26%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The service decision assumes, without calibration data, that after a cloud redetection the confidence score for a previously unknown object is uniformly spread around the query threshold; if actual large-model confidence is centered elsewhere, the accept/reject rule is not grounded.","fun_headline_variants_meta":{"raw":{"variants":["LVM-guided MPC saves 26% trip time in open driving","Simulation-timed cloud calls speed up autonomous driving","Collaborative planning: local MPC + cloud LVM, 26% faster","Opportunistic cloud queries cut navigation time by 26%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1423,"prompt_tokens":944,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":560,"tokens_out":479,"duration_ms":4646,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:26:16.276600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a test set of images with unknown objects, run the cloud detector on objects the local model flags as low-confidence, and compare the resulting confidence distribution to the assumed uniform band around $C_{\\text{threshold}}$; also compare CFS's accept/reject decisions against whether cloud guidance actually shortens the MPC trajectory in the simulator.","supporting_citations":[{"cited_title":"Rilaas: Robot inference and learning as a service,","cited_arxiv_id":null,"evidence_quote":"Defines the periodical cloud collaboration baseline (PCS) that OCP is compared against."},{"cited_title":"Multi-uncertainty aware autonomous cooperative planning,","cited_arxiv_id":null,"evidence_quote":"Provides the uncertainty-aware inflation distance formulation reused in the collision-avoidance constraint."},{"cited_title":"EPSILON: An efficient planning system for automated vehicles in highly interactive environ- ments,","cited_arxiv_id":null,"evidence_quote":"Provides forward simulation for interactive planning, the basis of CFS's trajectory improvement estimate."},{"cited_title":"Planning for robotic exploration based on forward simulation,","cited_arxiv_id":null,"evidence_quote":"Provides forward simulation for exploration planning, cited as the basis for CFS's simulation approach."}],"review_version":1}