{"id":"edd13651-3b43-4f1e-986d-238f26ec6cc9","arxiv_id":"2506.08851","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SICNav, a bilevel optimization crowd navigation planner previously validated in simulation and lab, was deployed in real indoor and outdoor pedestrian environments and ran for roughly two hours with acceptable solver speeds.","lead":"This workshop paper reports the first field deployment of SICNav, a bilevel MPC crowd navigation planner, on a Clearpath Jackal robot across indoor and outdoor areas of the TUM campus. The robot logged 6.73 km over about 1 hour 51 minutes, with preliminary analysis of solve times and manual takeovers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 6.73 km 'autonomous' distance may include manual takeover segments; the paper never states how takeovers are excluded, so the central deployment claim is not yet verifiable.","rationale":"The reader identified the ORCA human-model assumption as the weakest point, and that is a legitimate concern about safety guarantees. However, it is a modeling issue that the paper itself partly acknowledges (e.g., 'treat any additional humans as static obstacles') and that is somewhat secondary for a preliminary field-deployment workshop paper. The more directly load-bearing issue is the integrity of the central quantitative claim: 6.73 km of autonomous navigation. If manual takeover segments are included in that distance, the paper's headline result is not what it appears to be. This is a concrete, checkable accounting problem. The reader's verdict was already CONDITIONAL, and the conditions (release data, clarify metrics) would cover this concern, so no change to the verdict is needed. However, the reader's stated weakest assumption is not the same as the one identified here, hence 'disagree' on agreement_with_reader.","tokens_in":9837,"tokens_out":4006,"duration_ms":44462,"concrete_test":"Request the recorded rosbag logs (or a per-session table) containing timestamps for autonomy-active states, manual takeover intervals, and odometry. Recompute the total distance and duration only while the SICNav controller is issuing velocity commands and no manual takeover is active. Compare the recomputed values to 6.73 km and 1h51m; if the autonomous-only distance is materially lower (e.g., more than 5% difference) or takeover duration is non-negligible, the deployment claim should be revised. Alternatively, analyze the linked video to identify takeover segments and measure their duration and distance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest empirical claim is that the robot 'navigated autonomously for nearly one hour and 51 minutes, traversing 6.73 km' (Sec. III-B, Table I). The same table reports 49 manual takeovers. The paper never defines whether Duration and Distance are computed only while the SICNav controller is in command, or whether they include time and distance during manual takeover and recovery. The reported takeover frequency is computed as #TO / Duration (e.g., sidewalk: 27 / 2224 s = 0.0121 s^-1), which is consistent with Duration being total wall-clock time including manual segments, not purely autonomous time. If manual segments are included, the headline '6.73 km of autonomous navigation' is inflated and unsupported. The paper's own limitation section acknowledges that some takeovers were due to nondrivable areas, and during those takeovers the robot is not operating under SICNav. This is a direct accounting/definition problem for the central claim, not a theoretical model mismatch. Without a clear autonomy-active flag and per-segment logs, the number 6.73 km cannot be interpreted as autonomous distance.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an initial field deployment of SICNav, a bilevel MPC planner that couples robot trajectory planning with ORCA-based predictions of human motion, on a Clearpath Jackal robot. The system integrates Cartographer 2D SLAM, YOLOv9/aUToTrack-based human detection and tracking, costmap-based obstacle representation, and hybrid A* global planning. The authors report 6.73 km of autonomous navigation over 1h51m in previously unseen indoor, sidewalk, and campus environments, along with wall-clock statistics on manual takeovers, detected human counts, and SICNav solve times. They also present a preliminary simulation study comparing SICNav, SICNav-Diffusion, and a non-interactive CVMM MPC under additive state-estimation noise. The paper is explicitly framed as preliminary, with a stated plan for more extensive evaluation and comparison in a final version.","tokens_in":10017,"tokens_out":4492,"duration_ms":52679,"significance":"If the reported field statistics are verifiable, this is a useful systems-integration milestone: it demonstrates that a bilevel interactive MPC crowd-navigation algorithm can run in real time on a mobile robot in previously unseen environments with noisy localization and perception, which goes beyond the earlier in-lab VICON-based experiments. The paper does not claim new theoretical results; its value is the deployment architecture, the adaptation of aUToTrack to lidar intensity/range images, and the preliminary field data. The simulation robustness comparison is a reasonable first step, but the current evidence is too thin to substantiate the headline autonomy claims, and the safety/interactivity claims are stated without acknowledging the model assumptions under which they hold.","major_comments":[{"comment":"The headline claim that the robot 'navigated autonomously for nearly one hour and 51 minutes, traversing 6.73 km' is not verifiable as stated because the paper never defines how Duration and Distance are computed with respect to the 49 manual takeovers. If these quantities include wall-clock time and distance while the operator is in control or during takeover recovery, then the autonomous distance is inflated. The authors must state explicitly that Duration and Distance are computed only while the SICNav controller is in command, define an autonomy-active flag, and provide per-segment logs that separate autonomous, takeover, and recovery intervals. Without this, the central deployment claim cannot be interpreted.","section":"Sec. III-B, Table I"},{"comment":"The real-time performance claim is not supported by the reported statistics. The paper states that SICNav re-plans at 10 Hz, which corresponds to a 0.1 s deadline, yet the mean solve times are 0.100-0.114 s and are compared only to the 0.25 s MPC discretization period. The histograms also show a tail of larger solve times. The authors should report the maximum and 99th percentile solve times, the number of deadline misses, and whether the replanning loop is synchronous, asynchronous, or pipelined with the previous solution. As written, 'acceptable solve times' is a claim about the mean, not about whether the controller actually met its real-time schedule.","section":"Sec. III-B, Fig. 7"},{"comment":"The statements that SICNav is 'collision-free by construction' and ensures safety are model-level claims, not field-validated guarantees. In the deployment, these claims rely on the assumptions that ORCA accurately models human reactions, that the intended velocity is well estimated by projecting the latest observed velocity forward with a fixed horizon, and that modeling only the two humans with lowest time-to-collision with ORCA while treating all other humans as static is sufficient in real crowds. The paper should explicitly scope the safety claim to these model assumptions and should report safety-relevant events from the field (near-misses, collisions, interventions taken because of unexpected human behavior) so that readers can assess whether the assumptions actually held in the tested environments.","section":"Sec. I and Sec. II-E.3"}],"minor_comments":[{"comment":"The text says noise is sampled from a zero-mean Gaussian with standard deviation sigma in {0, 0.2}, but the discussion refers to performance changing as sigma increases 'beyond 0.03' and the plots show many intermediate levels; please specify the full set of noise levels and how the 200 scenarios are distributed across them.","section":"Sec. III-C.1, Fig. 8"},{"comment":"The percentile rejection thresholds (5th and 65th) used to filter lidar points from detection contours are introduced without justification or sensitivity analysis; please state how these values were selected and whether the tracking results are sensitive to them.","section":"Sec. II-C"},{"comment":"The intended human velocity is estimated 'by projecting the agent's latest observed velocity forward with a fixed time horizon,' but the horizon length is never specified; please give its value and comment on its effect on the results.","section":"Sec. II-E.2"},{"comment":"There is a typo in the caption ('Point cCloud' should be 'Point Cloud'), and the detection model name is written inconsistently as 'YOLOv9' and 'Yolov9' in the text.","section":"Fig. 3 caption"},{"comment":"The abstract and conclusion say 'nearly 7 km' and 'nearly two hours,' while Table I provides exact values (6.73 km, 1:51:41); please either cite exact numbers consistently or clarify that the rounded values are approximate.","section":"Abstract and Sec. IV"}],"recommendation":"major_revision","confidential_remarks":"This is a workshop-style preliminary deployment report. The main risk is the unverifiable accounting of the headline autonomous distance, which must be fixed before the central claim can be accepted. I would also ask the editor to consider whether the venue expects baseline field comparisons; the authors already acknowledge their absence, so I do not treat that as a fatal flaw, but the paper should be positioned as a preliminary systems report rather than as a comparative evaluation. The heavy self-citation is understandable given that SICNav is the authors' own algorithm, though the novelty relative to the prior T-RO paper should be made explicit in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a straightforward field-deployment report for a planner that already appeared in T-RO [9]. The useful new content is the systems integration and the preliminary field data; the main caveat is that the headline '6.73 km autonomous' is under-defined.\n\nThe systems description is concrete and credible: sensor preprocessing with Patchwork++, Cartographer SLAM, the lidar-intensity/reflectivity adaptation of aUToTrack (which sensibly removes the camera-to-lidar calibration problem), and the SICNav bilevel MPC formulation. The aggregate statistics—solve times around 0.10–0.11 s versus a 0.25 s control period, takeover frequencies per environment—are internally consistent and are presented without overclaiming. The simulation noise study is preliminary but a reasonable first step toward comparing SICNav against baselines under degraded state estimation. The authors are also upfront about limitations and about what they plan for a final version.\n\nThe soft spots are real but fixable. Most important is the accounting of autonomous distance. Table I reports 49 manual takeovers, yet the paper never states whether Duration and Distance include time/distance during operator control. The takeover frequency is computed as #TO / Duration, which suggests Duration is total wall-clock time including manual segments. If so, '6.73 km of autonomous navigation' is inflated and not verifiable until the authors provide an autonomy-active flag and per-segment logs. Second, no raw data or code are released, so none of the aggregate numbers are independently checkable. Third, safety is assessed only via manual takeover counts; that is thin, though the authors acknowledge it. Fourth, the noise simulation text says σ∈{0,0.2} but the figure axis shows a continuum; the sampling procedure needs clarification.\n\nFor a workshop paper this is a reasonable preliminary report, and the field deployment itself is valuable—real-world deployments of interactive bilevel MPC are rare. For a journal version, the distance-accounting issue would be load-bearing. As a workshop contribution, I would send it to peer review rather than desk reject, but the referee should insist on defining the autonomous intervals and ideally releasing the logs. It is a conditional accept, not a reject.","headline":"Deployment report of a published planner; the useful contribution is the field integration and data, but the headline distance claim needs a clearer definition of 'autonomous' before the numbers can be trusted.","tokens_in":10593,"tokens_out":1856,"would_cite":false,"duration_ms":21992,"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":"SICNav, a crowd-navigation planner that couples robot and human motion in one bilevel optimization, completed 6.73 km of autonomous field navigation in roughly two hours under real noisy sensing.","keywords":["crowd navigation","bilevel optimization","model predictive control","ORCA","field deployment","lidar perception","human tracking","autonomous navigation"],"falsifier":"Run the deployed system in a corridor where three or more people approach from different directions with crossing paths, and record all robot-human distances and manual takeovers; a collision, a repeated freeze, or a takeover caused by the two-humans-only model would show that the field guarantee does not hold outside the modeled regime. A simpler quantitative check is to monitor MPC solve times: any recorded solve time exceeding the 0.25 s control period would falsify the real-time claim.","tokens_in":9615,"feed_emoji":"🤖","tokens_out":10027,"duration_ms":101371,"temperature":0.7,"pith_summary":"This paper reports the first field deployment of SICNav, a crowd-navigation approach that folds human motion prediction and robot planning into a single bilevel nonlinear MPC problem. The claim is that the method can drive a small ground robot through previously unseen indoor and outdoor pedestrian spaces using only lidar, IMU, and wheel odometry, without motion capture. Over 1h51m of autonomous operation the robot traveled 6.73 km, with mean optimization solve times of 0.10 to 0.11 s, comfortably below the 0.25 s control period. If this holds, interactive crowd navigation is not limited to lab settings with clean state estimates; it can run in real time on noisy, real-world perception.","feed_headline":"Robot navigates 6.73 km through live crowds with a 0.1 s replan loop","feed_subtitle":"Bilevel MPC that predicts human reactions solves in about 0.1 s on a Jackal in previously unvisited spaces.","key_machinery":"The load-bearing component is the SICNav bilevel MPC: a nonlinear program whose outer level chooses the robot's trajectory while the inner level solves, for each human, an ORCA reciprocal-collision-avoidance problem; the inner problems are replaced by their Karush-Kuhn-Tucker optimality conditions so the whole system is optimized jointly. Around this, the field system adds camera-free human perception by running YOLOv9 segmentation on lidar intensity and reflectivity images, tracking detections with Kalman filters, and filtering tracked humans out of the static obstacle map so the optimizer can treat them interactively rather than as walls.","core_discovery":"On the paper's own terms, the central discovery is that a planner which treats human reactions as part of the optimization, replacing each human prediction subproblem with its optimality conditions and solving the combined problem, remains computationally tractable and behaviorally sensible when embedded in a full autonomy stack with SLAM, lidar-based human detection and tracking, and costmap-based static obstacles. The deployed system models the two humans with the lowest time-to-collision as interactive agents using ORCA and treats everyone else as static obstacles; this keeps the nonlinear program small enough to solve at 10 Hz. The authors report 6.73 km traversed and call the results preliminary. They present this as evidence that collision-free-by-construction interactive planning can run outside a controlled laboratory.","pith_inferences":["A natural stress test is to raise the number of interactively modeled humans from two to three or four as solver hardware improves; the paper's two-human cap is a computational simplification, not a modeling necessity, so the same bilevel structure has a clear scaling path.","Because the perception pipeline runs on lidar intensity images rather than a camera, it may remain functional in low-light or rain, where RGB cameras degrade; this is an untested implication of the camera-free design.","If the authors perturb localization, detection, and tracking separately and add perception delays in simulation, they may find that one component, not the planner, sets the ceiling on field performance; their current aggregate-noise simulation cannot isolate this yet.","Manual-takeover frequency could itself be used as a deployment metric: comparing takeover rates across methods in identical environments would quantify how much interactive modeling helps under realistic noise."],"forward_implications":["SICNav can act as a drop-in replacement for the local planner in a standard ROS navigation stack, since it consumes the same costmap and velocity-command interfaces.","The measured solve times of about 0.1 s are sufficient for 10 Hz replanning with a 2 s horizon, matching the control loops of small ground robots.","Filtering the two most imminent humans out of the costmap and modeling them as interactive prevents the optimization from becoming intractable while keeping behavior responsive.","In the authors' simulation noise study, SICNav's advantage over a constant-velocity non-interactive MPC narrows once the standard deviation of state-estimation noise grows past 0.03, indicating that perception quality sets a floor on interactive planning gains."],"supporting_citations":[{"why":"Defines the SICNav bilevel MPC formulation and reports prior lab-robot validations that this deployment builds on.","marker":"[9]"},{"why":"Provides the ORCA reciprocal collision avoidance model used as the human motion predictor and interaction constraint.","marker":"[10]"},{"why":"Supplies the 2D lidar SLAM system used to localize the robot and build maps of previously unseen areas.","marker":"[11]"},{"why":"Provides the YOLOv9 segmentation network run on lidar intensity and reflectivity images for human detection.","marker":"[12]"},{"why":"Supplies the aUToTrack detection-tracking and data-association method adapted for human identity maintenance.","marker":"[13]"},{"why":"Provides the Patchwork++ ground segmentation that preprocesses lidar point clouds for mapping and obstacle detection.","marker":"[14]"},{"why":"Supplies the acados solver used to solve the KKT-reformulated bilevel MPC in receding horizon at 10 Hz.","marker":"[16]"}],"fun_headline_variants":["SICNav completes 6.73 km field run, replanning every 0.1 s","Bilevel MPC handles human-robot interaction for 6.73 km of autonomy","Interactive MPC predicts human moves, robot sails through crowds","Field-deployed SICNav: bilevel optimization plans with human reactions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The safety and interactivity guarantees collapse if human motion is not well described by ORCA with intent guessed from the latest observed velocity, or if a real crowd has more than two humans approaching simultaneously, since additional humans are treated as static obstacles.","fun_headline_variants_meta":{"raw":{"variants":["SICNav completes 6.73 km field run, replanning every 0.1 s","Bilevel MPC handles human-robot interaction for 6.73 km of autonomy","Interactive MPC predicts human moves, robot sails through crowds","Field-deployed SICNav: bilevel optimization plans with human reactions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000723,"raw_usage":{"total_tokens":3207,"prompt_tokens":873,"completion_tokens":2334,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2250}},"tokens_in":489,"tokens_out":2334,"duration_ms":23454,"temperature":1.0,"reasoning_tokens":2250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:00:02.392994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the deployed system in a corridor where three or more people approach from different directions with crossing paths, and record all robot-human distances and manual takeovers; a collision, a repeated freeze, or a takeover caused by the two-humans-only model would show that the field guarantee does not hold outside the modeled regime. A simpler quantitative check is to monitor MPC solve times: any recorded solve time exceeding the 0.25 s control period would falsify the real-time claim.","supporting_citations":[{"cited_title":"SICNav: Safe and Interactive Crowd Navigation Using Model Predictive Control and Bilevel Optimization,","cited_arxiv_id":null,"evidence_quote":"Defines the SICNav bilevel MPC formulation and reports prior lab-robot validations that this deployment builds on."},{"cited_title":"Real-time loop closure in 2d lidar slam,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D lidar SLAM system used to localize the robot and build maps of previously unseen areas."},{"cited_title":"aUToTrack : A Lightweight Object Detection and Tracking System for the SAE AutoDrive Challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies the aUToTrack detection-tracking and data-association method adapted for human identity maintenance."}],"review_version":1}