{"id":"2a778cf1-3a03-4cfc-8d48-4833aebfc1d4","arxiv_id":"2504.19399","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A leader-following system stores leader appearance features by distance and switches between chasing, following, planning, and retreating states, which improved simulated follow success from below 22% to 96.9%.","lead":"A new robot control framework uses segmentation instead of object detection, so a robot can follow anything from a person to a drone and can keep tracking after the leader leaves the camera view. It adds a distance-based memory buffer and a goal-aware planner, and reports much higher follow success than existing methods in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The distance frame buffer's per-bin confidence selection is unvalidated; partial-view embeddings can still dominate, and centroid-based distance binning is biased under partial visibility.","rationale":"The reader's weakest assumption correctly identifies confidence-score reliability as the load-bearing premise, and I agree that no validation is provided. I add a more specific technical concern: even if S were well-calibrated, the bin assignment in Eq. (2) depends on a centroid computed from a partial point set, so corrupted embeddings can be stored in the wrong bin under the exact failure mode the framework targets. The conditional verdict is appropriate because the concern is addressable with targeted instrumentation and does not by itself disprove the central claim; the ablation table provides some support, but the load-bearing mechanism remains unverified. The paper's lack of a code URL and undocumented baseline tuning further limit independent checking, but the proposed logging/ground-truth-selection test would settle whether the distance frame buffer works as claimed. Thus I do not recommend changing the reader's CONDITIONAL verdict, but I would require this validation as a condition for acceptance.","tokens_in":12666,"tokens_out":5209,"duration_ms":58736,"concrete_test":"Instrument the playground simulation to log, for every timestep and every candidate embedding, the confidence score S(η_t), the mask IoU against a full-body ground-truth mask, the true distance to the leader, and the bin selected by Eq. (2). Then check whether any stored embedding in the distance frame buffer has low IoU (e.g., IoU < 0.5) or is assigned to the wrong distance bin due to the biased centroid. If a substantial fraction of stored embeddings are partial-view or mis-binned, the distance frame buffer premise fails. A complementary intervention: rerun the same scenario but replace each bin's stored embedding with the highest-IoU full-body embedding from that bin; if follow success changes materially, confidence-based selection is the bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism for re-identifying a leader after FOV loss is the distance frame buffer of Eq. (2), which keeps, for each distance bin, the historical embedding with the highest confidence score S(η_t). The paper's own motivation (Sec. I) states that just before leaving the FOV the leader yields 'partial and low-quality features' that 'can still dominate the bootstrapping process.' Replacing a global buffer with per-distance bins only helps if, within each bin, the arg-max-S embedding is actually a clean full-body view. No calibration or validation of S is provided, so a high-confidence partial-view embedding can still be stored and later used for re-identification. Additionally, the bin index in Eq. (2) uses d(η_t) = |p̄_l^t − p_f^t|, where p̄_l^t is the centroid of the visible leader points; when the leader is partially visible near the FOV edge, this centroid is biased, so the corrupted embedding can be stored in the wrong distance bin. The ablation FE vs FE-N-DFB (96.9% vs 62.5% follow success) is the only quantitative evidence for the buffer's benefit, but it does not isolate whether per-bin confidence selection is reliable; it could be confounded by other differences such as the number of retained embeddings or interactions with the goal-aware states.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified leader-following framework that combines a SAM2-based segmentation module with two memory mechanisms (a temporal memory buffer and a newly introduced distance frame buffer), a goal-aware state machine (chasing, following, planning, retreating, switching), and a graph-based trajectory planner. The method is evaluated in Gazebo using replayed leader trajectories across four scenarios with two external baselines (Alaa and SA-MPC) and two ablations, plus qualitative real-world demonstrations with a legged robot. The headline simulation results are a 96.9% follow success rate, a 10.7% average leader loss time ratio, a 1.8% collision rate, and a 2.0 m average distance, reported as best among all compared methods.","tokens_in":12996,"tokens_out":6247,"duration_ms":67666,"significance":"If the empirical results hold, the paper provides a useful integration for open-set leader following with reacquisition after the leader leaves the field of view. The simulation protocol, which replays recorded leader trajectories multiple times, is reasonable, and the ablation table indicates that both the distance frame buffer and the graph-based planner contribute to the reported improvement. The breadth of leader types (human, ground robot, UAV, legged robot, stop sign) is a strength, and the promise of released code and video is valuable. The work is primarily an integration of existing perception and planning components with a new distance-conditional memory design; its novelty is moderate. The main weaknesses are in evaluation rigor: inconsistent abstract comparisons, missing variance and baseline-tuning details, and an unvalidated confidence-selection mechanism in the distance frame buffer.","major_comments":[{"comment":"The four improvement percentages in the abstract are not computed against a single baseline. Specifically, 75.1% is the follow-success gain over Alaa (96.9 - 21.8), 13.1% is the loss-time reduction over Alaa (23.8 - 10.7), 65.1% is the collision-rate reduction over Alaa (66.9 - 1.8), and 0.4 m is the distance reduction over SA-MPC (2.4 - 2.0). Presenting these together as 'improvements' obscures the fact that the reference method changes from metric to metric. Please either report each percentage with its explicit baseline or use one consistent baseline throughout. Additionally, the table would be much stronger with per-scenario results and variance information.","section":"Abstract and Table I"},{"comment":"The distance frame buffer stores, for each distance bin, the historical embedding with the highest confidence score S(eta_t), but the paper neither defines S nor validates that a high confidence score corresponds to a clean, full-body view of the leader. This is load-bearing because the Introduction argues that partial FOV-edge features are low-quality yet can dominate bootstrapping; if such a corrupted embedding receives a high confidence score and is stored in a distance bin, the buffer may reproduce the same failure it is designed to avoid. The ablation FE-N-DFB shows that adding the buffer helps in aggregate, but it does not isolate the confidence-selection rule from other differences such as buffer capacity or retention policy. Please add an experiment that simulates partial visibility near the FOV edge and checks whether the selected embeddings are in fact clean full-body views, or compare confidence-based selection against random selection within each bin.","section":"Section II-B, Eq. (2)"},{"comment":"The real-world experiments are entirely qualitative: each scenario is illustrated by a sequence of images, with no measured follow success rate, leader loss time, collision events, average distance, or number of trials. Since the abstract claims that simulations and real-world experiments demonstrate competitive improvements, the absence of quantitative real-world results leaves a large part of the claim unsupported. Please add a quantitative real-world evaluation (even a small table with per-scenario metrics) or explicitly relabel these as demonstrations rather than experiments supporting the performance claims.","section":"Section IV"},{"comment":"The simulation results are reported as point estimates without confidence intervals or standard deviations, even though the protocol replays 10 leader trajectories four times per scenario. This makes it difficult to judge whether the differences between Follow Everything and the baselines are statistically meaningful or dominated by trajectory-level variation. The paper also gives no details on how the baseline hyperparameters were tuned (e.g., PID gains for Alaa, MPC weights and horizons for SA-MPC); without a documented tuning protocol, a reader cannot rule out that the baselines are under-tuned. Please provide per-scenario statistics and a short description of how each baseline was configured.","section":"Section III, Table I"}],"minor_comments":[{"comment":"The notation for the memory-buffer update is mathematically unclear: expressions such as 'arg max over forall eta_t cap eta_t notin [eta_1,...,eta_{i-1}]' are not standard set notation and do not define how the buffer is updated over time. Please rewrite these equations with explicit time indices and set definitions, and define where S(eta_t) comes from.","section":"Section II-B, Eqs. (1)-(2)"},{"comment":"The bar chart shows only means, while the text claims the method has the 'smallest variance' in average distance. Please add error bars or a violin/box plot so that the variance claim is visible and checkable.","section":"Figure 7"},{"comment":"The title given in the first abstract ('Follow Everything: A Leader-Following and Obstacle Avoidance Framework with Goal-Aware Adaptation') differs from the title in the full text ('Follow Everything: Goal-Aware Adaptation and Graph-Based Planning Towards Arbitrary Leader Following'). The manuscript should use a single consistent title.","section":"Title and Abstract"},{"comment":"The text says a new leader is 'most commonly triggered by a large language model [21]', but reference [21] is a vision-language SAM variant, not a large language model. Please correct the citation or the description.","section":"Section II-D, Switching state"},{"comment":"The quantity f(tau*_{t-1}, tau) is called a 'similarity' but, as written, it can take values outside [0,1] depending on alpha and the number of obstacles. Please state the intended range or add a normalization that makes the similarity interpretation precise.","section":"Section II-C, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal and the core idea is plausible, but the evaluation needs to be substantially strengthened before publication: the abstract's mixed-baseline percentages should be corrected, the distance-frame-buffer confidence mechanism needs at least one targeted experiment, and the real-world claims need quantitative support or explicit downgrading. I do not see a fundamental correctness error that would justify rejection; the issues appear addressable within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before reading. First, this is an honest, incremental systems paper: the new bits are a distance-frame buffer for re-identifying a leader after losing it, and a goal-aware state machine that feeds constraints into a graph-based planner. Second, the simulation ablation supports the central claim, but the paper never validates the confidence score that the buffer relies on, and the abstract's improvement percentages are computed against different baselines per metric.\n\nWhat's genuinely new: Eq. (2) keeps, per distance bin, the historical embedding with highest confidence. That's a simple but sensible idea for leader-following, where the leader's appearance changes with distance and the last frames before FOV loss are partial. The goal-aware adaptation with chasing/following/planning/retreating states is a reasonable way to handle the interaction. The graph planner is mostly inherited from the authors' own prior TEB work, but the integration is clean.\n\nThe paper does real work: four simulation scenarios, 10 replayed leader trajectories each, four trials per trajectory, 160 tests total, with two baselines and two ablations. The ablations show each component helps: removing the distance buffer drops success from 96.9% to 62.5%, removing the graph planner drops it to 81.3%. That is meaningful evidence.\n\nSoft spots, in proportion. The buffer's design depends on S(eta_t) being a trustworthy quality score. The paper's own introduction says partial low-quality features can dominate bootstrapping; the per-bin argmax only fixes that if, within each bin, the highest-S embedding is actually clean. No calibration or validation of S is provided. The stress-test also notes the bin index uses the centroid of visible leader points, which is biased under partial visibility. That's a real concern, but not fatal: the distance bins are coarse, and the overall buffer still improves performance empirically. I'd want a referee to ask for an ablation that isolates per-bin selection from simply keeping more embeddings.\n\nOther soft spots: no baseline tuning details, no code or video URL in the text (the abstract promises a website), and the real-world results are qualitative. The abstract's numbers are computed against different baselines per metric - 75.1% success vs Alaa, 0.4 m distance vs SA-MPC - which reads as cherry-picking even if unintentional.\n\nWho this is for: robotics researchers working on human/robot following, especially on legged platforms. It deserves serious review: the idea is plausible, the experiments are reproducible in structure, and the flaws are fixable. I'd send it to peer review with a request for code, a calibration check on the confidence score, and an ablation that teases apart the buffer's components.","headline":"A solid, incremental leader-following paper whose distance-frame buffer is the real novelty; the ablation evidence is decent, but the confidence-score assumption and the abstract's mixed-baseline percentages need a referee's attention.","tokens_in":13505,"tokens_out":2747,"would_cite":false,"duration_ms":26917,"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":"Legged robot follows anything, even lost leaders, with 96.9% success","keywords":["leader following","legged robots","segmentation model","re-identification","distance frame buffer","graph-based planning","goal-aware adaptation","obstacle avoidance"],"falsifier":"In the playground scenario, record leader embeddings during partial occlusion near the field-of-view edge, then replace the distance-frame-buffer selection rule with \"store the most recent embedding per distance bin\" instead of \"highest confidence\"; if re-identification success after reappearance does not drop, then confidence-based selection is not load-bearing for the reported gains.","tokens_in":12482,"feed_emoji":"🤖","tokens_out":5428,"duration_ms":47216,"temperature":0.7,"pith_summary":"The paper proposes a single leader-following framework that replaces category detectors with a segmentation model, so the leader can be any object — a person, a wheeled robot, a quadrotor, or a stop sign. Its central claim is that a distance frame buffer, which stores the highest-confidence leader embedding for each distance range, prevents partial-view features captured just before the leader leaves the field of view from corrupting re-identification. Combined with a goal-aware state machine (chasing, following, planning, retreating, switching) and a graph-based time-optimal planner, the framework lets a legged robot follow arbitrary leaders and recover after temporarily losing sight of them. In simulation across four scenarios and 160 tests, it reports a 96.9% follow success rate, a 10.7% average leader-loss time ratio, a 1.8% collision rate, and a 2.0 m average leader–follower distance, outperforming both baselines and ablations on every metric.","feed_headline":"Legged robot follows anything, even lost leaders, with 96.9% success","feed_subtitle":"Distance-binned memory and goal-aware planning keep the target in mind even after it leaves the camera view.","key_machinery":"The central object is the distance frame buffer: an array of $n_2$ slots, where slot $i$ stores the leader embedding with the highest confidence score $S(\\eta_t)$ among all historical embeddings whose follower–leader distance falls in the interval $[(i-1)\\Delta d, i\\Delta d]$. It works alongside a temporal memory buffer to bias segmentation toward a trustworthy reference at the current distance. The second piece of machinery is the goal-aware adaptation that, for each state, defines goal sets (lines, circles, or a fixed point) and velocity constraints fed into a graph-based trajectory optimizer: obstacle clusters become nodes, shortest collision-free connections become edges, and the planner enumerates detour trajectories, prunes them by homotopy class, and solves a time-optimal optimization to select the fastest feasible path.","core_discovery":"The paper's discovery is that leader re-identification after visual loss is improved by partitioning the memory of leader appearances by distance. Instead of maintaining only a single temporal buffer of recent embeddings (which the paper argues can be dominated by low-quality, partial-view features from just before the leader exits the field of view), the distance frame buffer keeps one high-confidence embedding per equal distance interval between follower and leader. When the leader reappears, the per-distance memory supplies a more representative embedding for matching. The planning side contributes a state machine that sets different goal sets and speed constraints for each interaction mode, and a graph-based planner that generates homotopy-distinct candidate trajectories and optimizes the fastest one under those constraints. The net claim is that this combination yields substantially higher follow success and lower collision rates than the compared baselines in simulation.","pith_inferences":["The distance-binned memory idea is not specific to leader-following; it could apply to any long-term object re-identification task where appearance varies with distance, such as warehouse tracking or persistent person following with intermittent views.","The confidence-score selection is untested under miscalibration: if the segmentation model's scores are not calibrated, the buffer might store worse embeddings than a recency-based or coverage-based scheme; an ablation storing the newest embedding per bin would isolate this.","Real-world results are demonstrated only qualitatively (a few indoor and outdoor clips); the simulation metrics are the only quantitative claims, so transferring them to field conditions with varied lighting, crowds, or uneven terrain remains an open testable question."],"forward_implications":["A robot following a person or another robot can keep re-identifying the target after occlusions or after the target leaves the camera view, as long as a high-confidence embedding was stored at a similar distance.","Switching to a speed-maximizing \"planning\" state toward the leader's last known pose reduces the chance of permanently losing the leader, compared to stopping or replicating the leader's path.","The goal-line and goal-circle constraint formulation lets the planner slide the trajectory endpoint along feasible goal sets, producing time-optimal paths without requiring a fixed waypoint.","The same planner and state machine can handle qualitatively different interactions (far, near, lost, retreating, switched) without separate planning algorithms per scenario."],"supporting_citations":[{"why":"Provides the Alaa baseline that uses SAM2 segmentation with PID planning and a temporal buffer, the main comparison for re-identification.","marker":"[10]"},{"why":"Provides the SA-MPC baseline (YOLO11 detection with an MPC planner) against which the framework's perception and planning claims are compared.","marker":"[11]"},{"why":"Defines the four evaluation metrics (follow success rate, leader loss time ratio, collision rate, average distance) used in the experiments.","marker":"[13]"},{"why":"Supplies the text-prompted segmentation model whose embeddings and confidence scores the distance frame buffer relies on.","marker":"[21]"},{"why":"Defines the obstacle-group clustering and dynamic-obstacle constraints used inside the graph-based planner.","marker":"[22]"},{"why":"Defines the goal-line formulation that the chasing state extends into goal line sets.","marker":"[23]"},{"why":"Provides the homotopy-class framework used to prune redundant candidate trajectories.","marker":"[24]"},{"why":"Supplies search-based path planning with homotopy class constraints, used to deduplicate detour trajectories.","marker":"[25]"},{"why":"Provides the integrated time-optimal trajectory planning and optimization that underpins the graph-based planner.","marker":"[26]"}],"fun_headline_variants":["Distance-binned memory lets robots reacquire lost leaders","Goal-aware planner keeps robots on track when leader vanishes","Lost leader? Distance memory and adaptive planning recover it","Robots follow any leader via segmentation and distance-based memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the segmentation model's confidence score telling which stored embedding will best re-identify the leader later, even when the leader was only partly visible when that embedding was captured; if confidence is miscalibrated, the distance buffer may store the very features that cause re-identification failure.","fun_headline_variants_meta":{"raw":{"variants":["Distance-binned memory lets robots reacquire lost leaders","Goal-aware planner keeps robots on track when leader vanishes","Lost leader? Distance memory and adaptive planning recover it","Robots follow any leader via segmentation and distance-based memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000766,"raw_usage":{"total_tokens":3373,"prompt_tokens":896,"completion_tokens":2477,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2412}},"tokens_in":512,"tokens_out":2477,"duration_ms":16495,"temperature":1.0,"reasoning_tokens":2412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:54:06.188143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the playground scenario, record leader embeddings during partial occlusion near the field-of-view edge, then replace the distance-frame-buffer selection rule with \"store the most recent embedding per distance bin\" instead of \"highest confidence\"; if re-identification success after reappearance does not drop, then confidence-based selection is not load-bearing for the reported gains.","supporting_citations":[{"cited_title":"Safe and robust human follow- ing for mobile robots based on self-avoidance mpc in crowded corridor scenarios,","cited_arxiv_id":null,"evidence_quote":"Provides the SA-MPC baseline (YOLO11 detection with an MPC planner) against which the framework's perception and planning claims are compared."},{"cited_title":"Efficient motion planning based on kinodynamic model for quadruped robots following persons in confined spaces,","cited_arxiv_id":null,"evidence_quote":"Defines the four evaluation metrics (follow success rate, leader loss time ratio, collision rate, average distance) used in the experiments."},{"cited_title":"Stc-teb: Spatial- temporally complete trajectory generation based on incremental opti- mization,","cited_arxiv_id":null,"evidence_quote":"Defines the obstacle-group clustering and dynamic-obstacle constraints used inside the graph-based planner."},{"cited_title":"Ga-teb: Goal- adaptive framework for efficient navigation based on goal lines,","cited_arxiv_id":null,"evidence_quote":"Defines the goal-line formulation that the chasing state extends into goal line sets."},{"cited_title":"Topology-driven parallel trajectory optimization in dynamic environ- ments,","cited_arxiv_id":null,"evidence_quote":"Provides the homotopy-class framework used to prune redundant candidate trajectories."},{"cited_title":"Search-based path planning with homotopy class constraints,","cited_arxiv_id":null,"evidence_quote":"Supplies search-based path planning with homotopy class constraints, used to deduplicate detour trajectories."},{"cited_title":"Integrated online trajec- tory planning and optimization in distinctive topologies,","cited_arxiv_id":null,"evidence_quote":"Provides the integrated time-optimal trajectory planning and optimization that underpins the graph-based planner."}],"review_version":1}