{"id":"4c5b7445-6ddf-4d5f-88bf-95f5a9227c06","arxiv_id":"2412.00162","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A goal-conditioned diffusion planner combined with a high-order control barrier function safety filter for unsignalized intersections, where the reported near-zero errors are driven by goal leakage and the safety proof omits cross velocity terms.","lead":"This paper combines a diffusion-based trajectory generator with a safety filter called Dynamic High-Order Control Barrier Functions to plan left turns, right turns, and straight crossings at intersections without traffic lights. The accuracy and safety claims are weakened because the model is given the exact final position from the data, and the safety derivation appears to omit mixed velocity terms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 19 omits the mixed second-order Lie derivative term -4 v·v_obs, so the DHOCBF safety guarantee is not established and the same-direction experiments in Fig. 4 are not evidence of correctness.","rationale":"The paper's central claims are (i) DHOCBF provides a hard safety guarantee in dynamic environments and (ii) DSC-Diffuser delivers realistic multi-task trajectory planning. The empirical evaluation of (ii) is indeed weakened by goal conditioning: the terminal state is fed as input, making ADE/FDE near zero by construction, as the paper itself acknowledges when it uses the penultimate step as the FDE. However, the more fundamental problem is (i): the DHOCBF constraint is derived incorrectly. The missing mixed Lie derivative term is not a modeling assumption but an algebraic error in the derivation of Eq. 19. If Eq. 19 is wrong, the forward-invariance proof does not go through, and the 'hard constraint' safety guarantee is unsubstantiated. This matters more than the evaluation issue because even a perfect diffuser evaluation would not rescue the safety claim. The reader's weakest_assumption focused on goal conditioning; I partially agree with that concern, and the reader did mention the Lie derivative issue in the rationale, but did not make it the load-bearing point. I therefore maintain the REJECT verdict: the theoretical safety contribution is not sound as written, and the empirical results are confounded. The paper does have positive aspects: it uses a real-world dataset, provides a modular planning-plus-filter architecture, and the diffuser without goals appears to learn reasonable multi-task behavior. But as submitted, the central safety guarantee is not established.","tokens_in":28798,"tokens_out":7048,"duration_ms":52638,"concrete_test":"Independently re-derive Eq. 19 from Definition 4 using h = ||s - s_obs||^2 - d_safe^2, keeping both mixed terms L_f L_fobs h and L_fobs L_f h, and verify that the second derivative contains -4 v·v_obs. Then rerun the Fig. 4(a) same-direction experiment with the corrected constraint; if the QP becomes infeasible or produces a trajectory with h < 0, the omitted term is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.C derives the 2nd-HOCBF constraint for the dynamic safety function h(s_t, s_obs(t)) = ||s_t - s_obs(t)||^2 - d_safe(t)^2 via Eq. 19. The derivation is incomplete. The full second derivative along the joint (s, s_obs) dynamics, with u_obs = 0 but f_obs(s_obs) = [v_obs, 0]^T, is L_F^2 h = L_f^2 h + 2 L_f L_fobs h + L_fobs^2 h, and the mixed term L_f L_fobs h evaluates to -2 v·v_obs, giving 2||v||^2 + 2||v_obs||^2 - 4 v·v_obs = 2||v - v_obs||^2. Eq. 19 instead uses only L_f^2 h + L_fobs^2 h = 2||v||^2 + 2||v_obs||^2, dropping the cross term. Because the omitted term is negative when the ego vehicle and the obstacle move in the same direction, Eq. 19 is a weaker constraint than the HOCBF condition (16)/(18) it is intended to implement. The validity experiments in Fig. 4 specifically use same-direction motion (v_obs = 1 or 3 m/s), so the reported safety there cannot be credited to the stated DHOCBF forward-invariance guarantee. The central safety claim of the paper is therefore unsupported as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a safety-critical trajectory planning framework for signal-free intersections, combining a diffusion-model planner (DSC-Diffuser) with task labels and goal conditioning and a Dynamic High-Order Control Barrier Function (DHOCBF) safety filter. The planner is trained on Interaction dataset trajectories (left-turn, straight, right-turn) in the MA scene and tested in both the MA and GL scenes. The DHOCBF is intended to be a less-conservative extension of HOCBF for dynamic obstacles. The manuscript reports state-of-the-art ADE/FDE values around 1e-3 with SR=1, and includes ablation studies on guidance weight and noise.","tokens_in":29151,"tokens_out":3827,"duration_ms":31705,"significance":"The integration of diffusion-based trajectory generation with a hard safety filter is a timely direction, and the paper includes comparisons with several baselines and a cross-scene generalization test. However, the main technical contribution, DHOCBF, rests on an incomplete Lie-derivative derivation, and the reported near-zero displacement errors are inflated by conditioning on the ground-truth final position, which the paper itself acknowledges fixes FDE at zero. As a result, neither the safety guarantee nor the realism claims are supported as written. The paper does contain useful experimental comparisons and an ablation structure, but the core validation is not trustworthy in its current form.","major_comments":[{"comment":"The second-order HOCBF condition for h = ||s - s_obs||^2 - d_safe^2 is derived by taking only L_f^2 h and L_f_obs^2 h. The full derivative along the joint dynamics includes the mixed term 2 L_f L_f_obs h = -4 v·v_obs, so the exact L_F^2 h equals 2||v||^2 + 2||v_obs||^2 - 4 v·v_obs = 2||v - v_obs||^2. Omitting the cross term makes Eq. (19) a weaker constraint than Eq. (18) requires when the ego vehicle and the obstacle move in the same direction. Since Fig. 4 explicitly uses same-direction motion (v_obs = 1 and 3 m/s), the numerical safety shown there cannot be attributed to the DHOCBF forward-invariance guarantee. Additionally, the time derivative of the dynamic safe distance d_safe(t) is not accounted for; if d_safe(t) is state- or time-dependent, additional terms enter ẖ. The authors should either correct the derivation or explicitly state the assumptions under which the cross term and ḋ_safe vanish.","section":"Section III.C, Eqs. (19)-(20)"},{"comment":"The planning problem as formulated in Eq. (1) uses s_{t+H}, the final position from the expert trajectory, as a condition. Section IV.C states: \"Since the positions of the last step as the goals are used in the DSC-diffuser, the FDE of the algorithm's output is fixed at zero.\" Consequently, the ADE values of 1.756E-03 (Table IV) and 0.1333 (Table V) are not comparable to those of baselines that lack this privileged oracle information. The claim of generating \"realistic, stable, and generalizable policies\" (abstract) is therefore not established by these metrics. The evaluation should either withhold ground-truth endpoints (for example, predict goals from the context) or report errors only on the portion of the trajectory not determined by the goal, and compare against baselines under identical conditioning.","section":"Section III.A and Section IV.C"},{"comment":"The DHOCBF parameters beta1 and beta2 are described as \"found by repeated experiments\" and take different optimal values under different conditions, and the guidance weight w is tuned per scenario. The less-conservative comparison with HOCBF in Figs. 4, 5, and 7 is thus partly a consequence of this per-condition tuning; no procedure is given for selecting these parameters in a new scenario, and no sensitivity analysis is reported. This would be a presentation issue if Eq. (19) were correct, but combined with the derivation gap it means the reported advantage of DHOCBF is not currently supported by a valid safety filter.","section":"Section V.A"}],"minor_comments":[{"comment":"The loss expression has an unclosed bracket in the second expectation term; the notation should be cleaned.","section":"Eq. (7)"},{"comment":"The symbol beta is overloaded: it denotes the conditional-information dropout probability in Eq. (7) and the HOCBF class K gains in Section III.C, which is confusing.","section":"Section III.B and III.C"},{"comment":"The symbol O_{t'} is used without definition; it presumably denotes an observation, but it should be explicitly defined.","section":"Section III.A, Eq. (3)"},{"comment":"The caption contains the phrase \"Orange points is are the goals\"; this is a typo.","section":"Fig. 1 caption"},{"comment":"L_f^2 h is written as 2 v_t^2, but with vector notation this should be 2||v_t||^2; the subscripts vo_xt and v_obs are also used inconsistently.","section":"Section III.C, Eq. (20)"},{"comment":"The abstract contains the grammatical error \"reduce the conservatism\"; it should be \"reduces conservatism\".","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The skeptical assessment is accurate. The derivation error in Eq. (19) is not a minor typo; it changes the constraint and invalidates the stated safety guarantee. The goal-conditioning issue is acknowledged in the paper itself and undermines the planning-quality claims. Both issues would require substantial rework and re-evaluation, so I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the qualitative idea here — a diffuser planner with a CBF-style safety filter for multi-task intersection driving — is reasonable, and the generalization experiment to an unseen intersection (GL) is the right instinct. But the paper's two headline claims rest on shaky ground. The DHOCBF derivation in Eq. 19 drops the mixed second-order Lie derivative, and the ADE/FDE numbers are largely an artifact of conditioning on the exact future endpoint.\n\nWhat is new and what works: The architecture of generating a reference trajectory with a conditional diffusion model and then minimally modifying it with a QP safety filter is clean and implementable. The authors test on the Interaction dataset, include an untrained scenario, and ablate goals and guidance weights. In the w/o-goals comparison, the safety filter does improve success rate in the GL scenario (0.878 to 1.0), which is evidence that the safety filter has some practical value even if the theory is not yet right.\n\nWhere it falls down: First, Eq. 19 is not the HOCBF condition for h(s_t, s_obs(t)) = ||Δ||^2 - d_safe^2. The full second derivative along the joint dynamics is 2||v||^2 + 2||v_obs||^2 - 4 v·v_obs. The paper keeps only the first two terms, so when ego and obstacle move in the same direction — exactly the case in Fig. 4 — the constraint used is weaker than the true HOCBF condition. The reduced conservatism relative to HOCBF is therefore expected, but the forward-invariance guarantee does not follow from the stated math. Second, the evaluation gives the model the true endpoint of the trajectory being predicted. The paper says this outright (\"the FDE of the algorithm's output is fixed at zero\"). With the oracle goal, ADE at 1e-3 and SR=1 are not meaningful evidence of multi-task planning quality; the baselines do not receive that information. Third, β1 and β2 are tuned by hand for each experiment, so the validity demonstrations are more like curated examples than robustness tests.\n\nMinor: the final PDF appears to contain a stray Chinese editing note in Figure 2 and several figure captions are garbled, which suggests the manuscript was not carefully prepared.\n\nThe citation pattern looks fine; refs [17] and [18] are the right anchors for dynamic CBF and HOCBF, and the paper is honest about its assumptions (non-reactive obstacles, perfect sensing) in the conclusion. Overall, this is an honest but incomplete paper. With a corrected derivation and an evaluation that either removes the oracle goal or gives the same goal information to all baselines, it could become a modest but useful contribution. As written, I would not rely on the safety guarantee.\n\nRecommendation: worth sending to peer review because the topic is active and the flaws are specific and fixable, but the editor should expect heavy revision and likely a different set of headline claims.","headline":"A well-intentioned diffuser-plus-CBF integration, but the DHOCBF derivation misses a mixed Lie derivative term and the evaluation feeds the oracle goal, so the main claims don't hold as written.","tokens_in":29686,"tokens_out":5444,"would_cite":false,"duration_ms":43770,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion-based planner conditioned on task and goal, with a dynamic high-order control barrier function as a safety filter, claims collision-free trajectories for left, straight, and right maneuvers at signal-free intersections.","keywords":["control barrier functions","diffusion models","trajectory planning","signal-free intersection","autonomous driving","multi-task learning","safety-critical control"],"falsifier":"Withhold the true final position from DSC-Diffuser at test time and measure ADE and FDE in the MA and GL scenes; if the errors rise toward the no-goal diffuser's values (ADE around 1), the $10^{-3}$-level precision is an artifact of the oracle goal rather than learned multi-task prediction. Separately, run the DHOCBF controller against an obstacle that brakes or accelerates after the planning step; if the distance to the obstacle ever drops below the safe threshold, the $u_{\\mathrm{obs}}=0$ assumption does not deliver the claimed safety guarantee.","tokens_in":28592,"feed_emoji":"🚗","tokens_out":8420,"duration_ms":65733,"temperature":0.7,"pith_summary":"This paper argues that one diffusion-based planner can learn all three maneuvers at an unsignalized intersection—left turn, straight, and right turn—when conditioned on the task label and the exit goal, and that its output can be made collision-safe by a Dynamic High-Order Control Barrier Function (DHOCBF) that accounts for the motion of surrounding vehicles. The planner, DSC-Diffuser, uses classifier-free task guidance and replaces the final state of each denoised trajectory with the specified goal, which drives displacement errors down to the order of $10^{-3}$ while the DHOCBF quadratic-program filter adjusts reference accelerations. On real-world unsignalized-intersection data, the method reports success rate 1 in both the trained scene and an unseen scene, and the DHOCBF is shown to track a reference trajectory more closely than standard HOCBF when obstacles move. The authors interpret these results as evidence for a safe, generalizable multi-task planning framework for signal-free intersections.","feed_headline":"One planner, three intersection maneuvers, zero collisions","feed_subtitle":"Diffusion planner plus safety filter handles left turns, straight, and right turns in seen and unseen intersections.","key_machinery":"The load-bearing object is the Dynamic High-Order Control Barrier Function. The barrier function is $h(s_t, s_{\\mathrm{obs}}(t)) = \\|p_t - p_{\\mathrm{obs}}(t)\\|^2 - d_{\\mathrm{safe}}(t)^2$, with a rectangle-based dynamic safe distance, and the second-order HOCBF constraint (Eq. 19) is expanded to include Lie derivatives $\\mathcal{L}_{f_{\\mathrm{obs}}} h$ and $\\mathcal{L}^2_{f_{\\mathrm{obs}}} h$ that encode the obstacle's velocity, so the safe set stays invariant when obstacles move. This constraint is enforced through a quadratic program that finds the control input closest to the diffuser's reference acceleration while satisfying the barrier condition and input limits. The other component is the goal-oriented, task-guided diffusion process: the denoiser is conditioned on the maneuver label via classifier-free guidance, and after each denoising step the final state-action pair of the sampled trajectory is overwritten with the given goal, which the paper reports as the main source of the low displacement errors.","core_discovery":"The paper's central claim is that combining a goal-conditioned diffusion planner with a dynamic control barrier function yields human-like, collision-free trajectories for left, straight, and right maneuvers at signal-free intersections, in a way that transfers to an intersection the model has never seen. The key step beyond prior work is formulating the barrier condition so that the obstacle's own motion enters the derivative constraint: instead of treating the other vehicle as a fixed obstacle, DHOCBF uses the second-order condition with Lie derivatives with respect to both the ego state and the obstacle state, making the safe set $C=\\{s : h(s,s_{\\mathrm{obs}})\\ge 0\\}$ forward invariant as the obstacle moves. The reported experiments show that, compared with HOCBF, DHOCBF keeps a smaller distance to a moving obstacle while still avoiding collision, and that DSC-Diffuser achieves success rate 1 in the MA (trained) and GL (untrained) scenes with average displacement error around $10^{-3}$ meters. The authors frame the result as a unified framework that recovers multi-task policies from expert demonstrations and enforces safety as a hard constraint without sacrificing efficiency.","pith_inferences":["The near-zero displacement errors are largely a measure of endpoint reconstruction: because the true final position of each recorded trajectory is supplied as the goal, the metric says how well the model reproduces a known ending, not how well it predicts an unknown one; a fairer accuracy test would condition on a sampled or predicted goal, and errors would likely rise to the no-goal level.","The safety guarantee rests on setting the obstacle's control input to zero, so DHOCBF is rigorous for obstacles that move predictably (constant velocity); for drivers who brake or swerve in response to the ego vehicle, a robust or game-theoretic extension would be needed, and the paper lists non-reactive surrounding vehicles as a limitation.","A direct extension suggested by the method is to make the class-K gains $\\beta_1,\\beta_2$ adapt online to the ego speed or the obstacle's heading, which could further reduce conservatism without violating the invariance condition; the authors name dynamic parameter selection as future work.","The goal-conditioned diffuser plus safety filter recipe should carry over to other structured maneuvers with known target lanes, such as roundabout entry or unprotected left turns at signalized intersections, where the exit position is also determined by the lane topology."],"forward_implications":["A single diffusion policy can represent all three intersection maneuvers, so a vehicle does not need separate planners for left turns, straight-through, and right turns.","DHOCBF works as a standalone safety filter that takes any reference control from any planner and minimally modifies it, which means the safety mechanism can be attached to planning methods other than the diffuser.","Because goal conditioning keeps displacement errors low in an untrained intersection, deployment in a new map with known exit lanes is plausible without retraining.","With the DHOCBF filter, the reported success rate is 1 in both the trained and the unseen scene, suggesting the integration does not trade away collision avoidance for trajectory accuracy."],"supporting_citations":[{"why":"Defines the HOCBF whose second-order barrier condition DHOCBF generalizes to moving obstacles.","marker":"[18]"},{"why":"Provides the INTERACTION dataset of real-world unsignalized intersection trajectories used for training and evaluation.","marker":"[19]"},{"why":"Supplies the denoising diffusion probabilistic model underlying the trajectory generator.","marker":"[48]"},{"why":"Provides classifier-free guidance, used to condition the diffuser on the maneuver label.","marker":"[49]"},{"why":"Introduces the diffuser planning formulation and the goal-constraint idea that the authors adapt.","marker":"[50]"},{"why":"Establishes the control-barrier-function quadratic program formulation the safety filter solves.","marker":"[51]"}],"fun_headline_variants":["Diffuser plus dynamic barrier: safe turns and straights","Safety filter learns all intersection moves from demos","Dynamic CBF + diffusion: collision-free at unsignaled intersections","Adaptive barrier keeps AVs safe in unseen intersections"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported near-zero errors assume the planner is handed the true final position of each recorded expert trajectory as a goal, while the safety guarantee assumes surrounding vehicles keep their current motion and do not react to the ego vehicle.","fun_headline_variants_meta":{"raw":{"variants":["Diffuser plus dynamic barrier: safe turns and straights","Safety filter learns all intersection moves from demos","Dynamic CBF + diffusion: collision-free at unsignaled intersections","Adaptive barrier keeps AVs safe in unseen intersections"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1560,"prompt_tokens":1052,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":668,"tokens_out":508,"duration_ms":4530,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:58:58.884573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Withhold the true final position from DSC-Diffuser at test time and measure ADE and FDE in the MA and GL scenes; if the errors rise toward the no-goal diffuser's values (ADE around 1), the $10^{-3}$-level precision is an artifact of the oracle goal rather than learned multi-task prediction. Separately, run the DHOCBF controller against an obstacle that brakes or accelerates after the planning step; if the distance to the obstacle ever drops below the safe threshold, the $u_{\\mathrm{obs}}=0$ assumption does not deliver the claimed safety guarantee.","supporting_citations":[{"cited_title":"Control barrier function based quadratic programs with application to adaptive cruise control","cited_arxiv_id":null,"evidence_quote":"Establishes the control-barrier-function quadratic program formulation the safety filter solves."}],"review_version":1}