acceptodds
Under review as a conference paper at ICLR 2027

ASAP: Arbitrary-Step Safe Planning with MeanFlow and Control Barrier Functions

Abstract

Generative models have emerged as a powerful paradigm for data-driven robotic planning. Existing flow-matching planners typically rely on multi-step sampling and lack principled mechanisms for enforcing safety constraints at test time. Bridging these limitations is essential for real-world deployment, where planners must support both fast generation and formal safety guarantees. We propose **A**rbitrary-Step **SA**fe **P**lanning (ASAP), a control-augmented MeanFlow framework for safe generation with any number of sampling steps, including one step. ASAP introduces a virtual control input governed by an affine barrier constraint that accounts for the sampling step size, yielding formal safety guarantees of generation for any number of steps. Without retraining, ASAP enforces unseen test-time constraints across navigation, locomotion, manipulation, and dexterous manipulation benchmarks, while planning at least faster than safety-guaranteed baselines on navigation, locomotion, and manipulation, and faster on dexterous manipulation. We further validate ASAP on a real-world robot handover task, where it maintains safety and high task performance under an obstacle, requiring only s per plan and achieving over a speedup over the safe baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Flow matching, Safety, Planning, Robotics, Generative Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QXDfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes ASAP (Arbitrary-Step Safe Planning), a framework that integrates MeanFlow with Control Barrier Functions (CBFs) to enable efficient and safe trajectory generation with an arbitrary number of sampling steps, including one-step generation. The key contribution is a discrete-time affine barrier constraint that enforces safety during the generative process without requiring model retraining.

Critical strengths
  • Efficient and Safe Motion Planning: The paper enables fast trajectory generation while providing formal safety guarantees, addressing an important limitation of existing generative motion planners.

  • Strong Theoretical Guarantees: The authors establish formal safety guarantees for arbitrary-step generation, including one-step sampling, and prove zero conservatism of the affine barrier constraints under the stated SDF assumptions.

  • Extensive Experimental Validation: The method is evaluated across diverse simulation benchmarks, including navigation, locomotion, and manipulation, as well as a real-world robotic handover task. The results demonstrate substantial improvements in planning efficiency while maintaining safety and competitive task performance.

Critical weaknesses

None

Questions for authors

Suggestions for Future Work: While the work presents strong theoretical and experimental results, I encourage the authors to consider evaluating the proposed framework in autonomous driving simulations. Given the safety-critical nature of autonomous driving, demonstrating the effectiveness of ASAP in such scenarios would further strengthen the practical relevance of the proposed safety guarantees and enhance the overall impact of the work.

4 comments
AuthorsOpenReview ↗

Response to Reviewer QXDf (1/1)

We sincerely thank the reviewer for the positive assessment of ASAP's efficiency, formal safety guarantees, and extensive validation across simulation and real-world experiments. We appreciate the recognition of our contributions and the valuable suggestion for future work.

Q1: Extension to autonomous driving.

Consider evaluating ASAP in autonomous driving simulations to further demonstrate its practical relevance in safety-critical applications.

We thank the reviewer for the positive assessment and valuable suggestion. We agree that autonomous driving is a promising application of ASAP. Recent studies [1,2] have demonstrated the effectiveness of flow matching for end-to-end autonomous driving, further motivating the application of our framework to this domain. In particular, ASAP provides a unified safety framework for both efficient one-step and flexible multi-step trajectory generation, enabling test-time safety correction without retraining. Our current evaluation focuses on validating these properties across navigation, locomotion, and robotic manipulation tasks. Extending ASAP to autonomous driving would require incorporating driving-specific constraints and evaluation protocols, which we consider an important direction for future work.

[1] Xing, Zebin, et al. "Goalflow: Goal-driven flow matching for multimodal trajectories generation in end-to-end autonomous driving." CVPR, 2025.

[2] Liu, Lin, et al. "Guideflow: Constraint-guided flow matching for planning in end-to-end autonomous driving." CVPR, 2026.

AuthorsOpenReview ↗

Additional Response to Reviewer QXDf: Autonomous Driving Evaluation

Motivated by the reviewer's suggestion, we conducted additional experiments on NAVSIM, an end-to-end autonomous-driving benchmark also used in recent flow-based planning studies [1,2].

Experimental setup. Following [2], we use its pretrained visual encoder to condition the generation of eight future trajectory waypoints. We formulate an SDF for roadside walls, where h≤0h\leq 0 indicates wall contact or collision, and evaluate ASAP across 20 driving scenarios against three baselines considered in our main paper: MeanFlow (MF), SafeFlowMatcher (SFM), and SafeDiffuser (S-Diff.) .

Results. As shown in the table below, ASAP achieves a 100100% safety rate across all 20 scenarios with one-step generation in only 0.4740.474 s in average. In contrast, MeanFlow is faster (0.0020.002 s) but achieves only 3030% safety. SFM achieves 5555% safety with 1616 steps and requires 128128 steps to reach 100100% safety, taking 6.0376.037 s (12.7×12.7\times slower than ASAP). Similarly, increasing S-Diff.'s sampling steps from 3232 to 256256 improves its safety rate from 4545% to 9595%, but still fails to achieve complete safety while increasing computation time from 3.5723.572 s to 22.54922.549 s. These results demonstrate ASAP's clear advantage in achieving both safety and computational efficiency for autonomous-driving trajectory planning.

Comparison of safety rate and planning time across 20 NAVSIM driving scenarios.

MethodSampling step(NN)Safety rate(↑\uparrow)Average time(↓\downarrow)
MF113030%0.0020.002
SFM16165555%1.0891.089
SFM128128100100%6.0376.037
S-Diff32324545%3.5723.572
S-Diff.64648585%6.3376.337
S-Diff.1281289595%11.71211.712
S-Diff.2562569595%22.54922.549
ASAP11100100%0.4740.474

We appreciate the reviewer's suggestion. These additional results further demonstrate ASAP's effectiveness in safety-critical autonomous-driving scenarios, strengthening the practical relevance of its safety guarantees and the broader impact of our work.

[1] Liu, Lin, et al. "Guideflow: Constraint-guided flow matching for planning in end-to-end autonomous driving." CVPR, 2026.

[2] Wang, Junli, et al. "Meanfuser: Fast one-step multi-modal trajectory generation and adaptive reconstruction via meanflow for end-to-end autonomous driving." CVPR, 2026.

Thank you for the authors detailed response and additional experiment. The NAVSIM experiment increases the practical relavance of the proposed approach. All my comments have been addressed

AuthorsOpenReview ↗

Author Response to Reviewer QXDf

We sincerely thank the reviewer for the positive feedback and for recognizing the additional autonomous-driving validation! We are glad that our revisions and experiments have fully addressed your comments. We would be grateful if you could consider whether the additional evidence changes your overall assessment. Of course, we fully respect your decision if you prefer to maintain your current rating.

Reviewer_YWCWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies test-time safety for generative trajectory planners. Existing safe diffusion/flow planners (SafeDiffuser, SafeFlowMatcher, SAD-Flower) enforce control barrier function (CBF) conditions in continuous time at every step of a multi-step sampler, which is slow, and their guarantees do not directly carry over to few-step samplers with large step sizes. The authors propose ASAP (Arbitrary-Step SAfe Planning), built on MeanFlow, which learns the average velocity over an arbitrary interval and therefore supports few-step and one-step generation.

Method. ASAP adds a control input u_{t_i} to each MeanFlow update (Eq. 8-9). At every step it requires the next sample to satisfy a discrete-time barrier condition h(tau_{t_{i+1}}) >= rho_{t_{i+1}} (Eq. 11), where rho is a relaxation schedule that tightens from -h_bar to 0 at the terminal step. This condition is linearized at the nominal update to give an affine constraint (Eq. 12), and the minimum-norm correction is obtained from a QP (Eq. 13). When h is an exact signed distance function (SDF), ||grad h|| = 1 and h varies linearly along its gradient, so the linearization is exact ("zero conservatism", Thm 4.2), the QP is always feasible (Remark 4.2, Thm B.1), and the terminal sample is safe for any number of steps N, including N = 1 (Thm 4.1, Cor 4.1). The appendix extends the guarantee to convex barriers (Thm B.2) and to barriers with bounded negative curvature via a robust margin (Thm B.3), and gives an O(1/N) bound on the approximation-induced conservatism (Thm B.4).

Validation. Experiments cover Maze2d (two to five unseen obstacles), Hopper (torso-height constraint), KUKA block stacking (tightened joint-range constraints), and Adroit Relocate (39-D state), plus a real Franka ice-cream handover task with an obstacle absent from the demonstrations. Baselines include MeanFlow, SafeFlowMatcher, SAD-Flower, Diffuser, classifier guidance, SafeDiffuser, and CoBL (plus FM, truncation, and rejection sampling in the appendix). The authors report zero constraint violation for ASAP on all tasks, task scores comparable to the best baselines, and planning at least 20x faster than safety-guaranteed baselines (about 5x on Adroit, over 200x in the real-world task, 0.04 s per plan). Ablations cover sampling steps, number/tightness of constraints, the relaxation parameter h_bar, and a comparison against direct post-optimization.

Critical strengths

S1. A practically relevant problem with consistent empirical results. Combining hard test-time safety with few-step/one-step generation targets a real bottleneck for deploying generative planners on robots, where multi-step safe samplers cost seconds per plan. Across Maze2d, Hopper, KUKA, and Adroit, ASAP reports zero violation of constraints unseen during training, keeps task scores close to the unconstrained MeanFlow model (Table 2, Table 6), and remains safe as the constraint is tightened (Fig. 3b) or the number of obstacles grows (Fig. 3c). The real-robot handover at about 0.04 s per plan (Table 3) shows the pipeline is deployable. If the comparisons hold up under the additional controls I request below, this is a useful engineering result for the community.

S2. Clear and honest theoretical treatment of the discrete-time setting. The paper states precisely why a continuous-time CBF condition at t_i does not imply safety at t_{i+1} for large steps, and replaces it with a discrete-time condition on the next sample. The exact-SDF case is worked out cleanly (Thm 4.1/4.2, proofs in B.2/B.4 are correct as far as I checked), and the authors go beyond it with the convex and bounded-curvature extensions and an explicit O(1/N) conservatism bound (Thm B.2-B.4), stating the assumptions under which each holds. This makes the scope of the guarantee easy to audit, which I value even though I have concerns about how much it adds (see Weaknesses).

Critical weaknesses

W1. The theoretical contribution is overstated: the "arbitrary-step safety guarantee" reduces to a final-step nearest-point projection, which is closely related to existing post-processing.

  • By the proof of Thm 4.1, terminal safety depends only on the last step, where rho_{t_N} = 0. With an exact SDF, the "QP" solution is the closed-form Euclidean projection u*Dt = (rho - h) grad h (Eq. 25/33). Hence safety for any N follows from projecting the final sample onto the safe set. The intermediate relaxation schedule rho_t, the linearization of Eq. 12, and the "zero conservatism" result (Thm 4.2) are not needed for the guarantee; they can only affect sample quality. Feasibility (Remark 4.2) is likewise trivial because u is unconstrained.
  • The introduction criticizes post-processing methods (Maze & Ahmed, 2023; Romer et al., 2024) for deviating from the learned distribution. Yet for N = 1, ASAP is exactly "one MeanFlow step followed by projection onto the boundary of S" (rho_1 = 0 places corrected states exactly on the boundary). The paper provides no analysis or measurement of distributional deviation, no comparison with the closest projection baseline, and no study of how a corrected, off-path state affects the next MeanFlow evaluation mu(tau, t, t+Dt), which was trained on on-path inputs.
  • "Conservatism" is defined only as the linearization residual (Eq. 14). This is not the quantity relevant to the stated objective (iii) in Eq. 2, i.e., minimal deviation from the expert distribution p_1. A zero linearization residual says nothing about how far the corrected sample is from the model's own distribution.
  • Consequently, the claim "first generative planner with formal safety guarantees for arbitrary-step generation, including the one-step case" and the novelty over SafeFlowMatcher/SAD-Flower/projection methods are not convincingly established. The real contribution looks closer to "MeanFlow + per-step SDF projection", which is useful but incremental.

W2. The empirical evidence does not isolate the benefit of the proposed safety mechanism, and the safety metric is largely tautological.

  • Missing key baseline / speed attribution. The 20x-200x speedups are measured against multi-step control-theoretic baselines (SafeDiffuser 17-69 s, SFM, SAD-F). Most of the gain comes from MeanFlow's few-step sampling, not from ASAP's safety design. The essential control is missing: the same MeanFlow model at the same N with no intermediate correction and a single closed-form SDF projection at the end. The only related ablation (Table 4, "post-opt") solves a quadratically constrained QP with a non-QP solver, which is a deliberately weak comparator. Moreover, ASAP's own overhead is 5-6x that of MF (0.06 s vs 0.01 s on Maze2d, 0.10 s vs 0.03 s on KUKA), even though the correction has a closed form for an SDF; this indicates that solver overhead, rather than the algorithm, dominates the measured time. Baseline times (e.g., SafeDiffuser 68.75 s on KUKA) may also reflect implementation cost rather than the method.
  • Safety metric is by construction. s-vio is computed with the same h on the same discrete waypoints that ASAP projects, so zero violation is guaranteed by the procedure and is not an empirical validation. What matters for robotics is not evaluated: (i) collisions between waypoints (a point mass can cross a thin obstacle between sparse waypoints; the guarantee covers only waypoints), (ii) dynamic feasibility/trackability of the corrected trajectories (the correction is applied independently per state, which breaks consistency between neighboring states; the conclusion acknowledges this as future work), and (iii) closed-loop executed safety (e.g., how Hopper actions are obtained from the corrected state trajectory is not explained). Zero safety margin (rho_1 = 0, states on the boundary) is also problematic under tracking/perception error, and the real-world setup assumes accurate motion-capture state.
  • Task-score evidence is weak. ASAP and MF have identical score on Maze2d (1.46 +/- 0.47) and Table 5 scores do not change at all with h_bar, suggesting the score is insensitive to the obstacle constraint, so "maintains task performance" is not demonstrated. Standard deviations are large (e.g., 1.75 +/- 0.83), and the number of seeds/episodes and any significance testing are not reported in the main text. The real-world study has 20 trials, compares only to MF/CG/S-Diff (not to SFM/SAD-F, the closest methods), and MF has s-vio 0.35 yet score 1.00, which needs explanation.
  • Scope of constraints. Constraints are simple (2-D discs, a scalar height threshold, joint-range boxes) for which exact SDFs and gradients are trivial. The theory relies on an exact differentiable SDF (||grad h|| = 1); general or learned constraints in high-dimensional configuration space, where exact SDFs are unavailable, are only covered by convex/bounded-curvature results without experiments. High-dimensional validation is a single Adroit task with a low-dimensional constraint. For multiple obstacles, h = min_j h_j is the exact SDF of the union only for non-overlapping obstacles (Appendix A), and ties on the medial axis are non-differentiable and handled only by remark.

Taken together, I think the paper has a useful engineering idea but the theory is thinner and the evidence less controlled than the framing suggests. Both issues look fixable with additional baselines, metrics, and a more careful statement of the contribution.

Questions for authors
  1. For an exact SDF, is the QP in Eq. 13 solved in closed form (Eq. 25/33)? If a generic QP solver is used, how much of ASAP's reported time (e.g., 0.06 s vs 0.01 s for MF on Maze2d) is solver overhead? How do Tables 2-3 change with the closed-form implementation?
  2. For N = 1, what is the difference between ASAP and "MeanFlow output followed by nearest-point projection onto the boundary of S"? Please report this baseline on all tasks of Table 2. This directly affects my assessment of novelty and of the speed claims.
  3. Since Thm 4.1 depends only on rho_{t_N} = 0, what is the actual role of the intermediate schedule rho_t and of h_bar (Eq. 10)? Is h_bar >= max{-h(tau_{t_0}), 0} taken per waypoint k or as a maximum over the horizon, and is it satisfied by the fixed values 1.0-3.0 in Table 5 for Gaussian initial samples far from S?
  4. Which constraint level of KUKA is used in Table 2, and how does it relate to Table 7 (30%) and Fig. 3b? How many seeds/episodes underlie each entry, and how are mean/std computed?
  5. Why is the Maze2d score identical for MF and ASAP (1.46 +/- 0.47) and invariant to h_bar (Table 5)? Is the score computed in an environment where the obstacles affect the outcome?
  6. In the real-world experiment, did MF (s-vio 0.35) cause physical collisions with the juicer while still scoring 1.00? How is the normalized collision margin measured on hardware, and how does the robot track the corrected trajectory?
  7. For multiple obstacles, how do you handle overlapping obstacles (where min_j h_j is not the exact union SDF) and points with non-unique nearest boundary? Do such points occur in your experiments?

Suggestions that could change my assessment (additional experiments)

  1. Evaluate continuous-time collision between waypoints, dynamic feasibility of corrected trajectories (e.g., displacement/velocity statistics, tracking error), and closed-loop executed safety.
  2. Quantify distributional deviation caused by the correction (e.g., distance to the unconstrained sample distribution, or the Wasserstein distance) and compare ASAP against final-step projection, rejection sampling, and the projection of Maze & Ahmed (2023) at matched N. Add SafeFlowMatcher/SAD-Flower (with matched step budgets or optimized implementations) to the real-world comparison.
  3. Include a safety margin (rho_1 > 0) analysis and robustness to perception noise.

Future work (does not affect my assessment)

  1. Extension to non-SDF barriers in realistic configuration-space collision checking, constraints coupling neighboring states (velocity/acceleration/dynamics), and input-constrained corrections.
7 comments
AuthorsOpenReview ↗

Response to Reviewer YWCW (1/5)

We appreciate the reviewer's recognition of ASAP's practical relevance in addressing the bottleneck of safe few-step planning, as well as the clear theoretical treatment of discrete-time safety. We also thank the reviewer for the constructive suggestions, which have motivated further experimental comparisons and clarifications addressed below.

W1+Q2: Novelty of ASAP versus final-only SDF projection.

Clarify how ASAP differs from final-only projection, particularly for one-step generation, and justify the design and benefits of intermediate safety corrections.

We would like to clarify that ASAP's main contribution is a unified framework for safety-guided generation with arbitrary sampling steps, integrating efficient SDF-based corrections with time-varying safety relaxation. We agree that, for (N=1N=1), ASAP is equivalent to MeanFlow followed by an SDF-based projection under the assumptions of Theorem 4.1. However, Fig. 3(a) shows that more complex tasks, such as KUKA, benefit from multiple sampling steps. For such tasks, ASAP goes beyond final-only projection by introducing a time-varying relaxation ρt\rho_t, motivated by prescribed-time safety control [1], which enables progressive safety guidance during generation. These intermediate corrections influence subsequent MeanFlow predictions, potentially avoiding substantial final corrections while preserving terminal safety.

To validate this distinction, we conducted additional ablations comparing ASAP with final-only SDF projection and full SDF projection at every step across all benchmarks, with complete results provided in the Table 7 and Appendix C.6. We highlight KUKA (N=4N=4) and real-world handover (N=2N=2) in the main paper as summarized in the table below, as these multi-step tasks better demonstrate the benefit of intermediate guidance. All three methods achieve zero safety violations, while ASAP obtains the highest task scores on KUKA (1.751.75 vs. 1.151.15/1.351.35 for Final-only/Full) and real-world handover (1.001.00 vs. 0.800.80/0.700.70). For one-step benchmarks (N=1N=1), ASAP and final-only SDF projection are mathematically equivalent under the exact-SDF assumptions, yielding equivalent safety and task performance.

Ablation of score result for SDF projection strategies

BenchmarkFinal-OnlyFullASAP
Kuka(N=4)1.15±0.791.15\pm0.791.35±0.961.35\pm0.961.75±0.831.75\pm0.83
Real robot(N=2)0.80±0.400.80\pm0.400.70±0.460.70\pm0.461.00±0.011.00\pm0.01

We also clarify that Theorem 4.2 establishes zero affine linearization error rather than zero distributional deviation. We have refined the corresponding claims to distinguish the efficiency of the SDF-based correction from preservation of the learned trajectory distribution.

[1] Song, Yongduan, et al. "Time-varying feedback for regulation of normal-form nonlinear systems in prescribed finite time." Automatica 83: 243-251, 2017.

AuthorsOpenReview ↗

Response to Reviewer YWCW (2/5)

W2-1+Q1: Speedup attribution and closed-form implementation.

Clarify whether ASAP's efficiency gains primarily come from MeanFlow or its safety formulation, and explain the QP solver overhead and potential speedup from closed-form corrections.

We agree that MeanFlow's few-step sampling contributes substantially to ASAP's efficiency. However, ASAP's key contribution is enabling safe few-step generation with formal guarantees without introducing expensive additional overhead. While MeanFlow provides fast generation, it does not guarantee safety. Existing safe planners, including SAD-Flower (SAD-F), SafeFlowMatcher (SFM), and SafeDiffuser (S-Diff.), generally rely on repeated safety corrections to achieve safety, introducing substantial computational overhead. Simply reducing these corrections to match MeanFlow's few-step generation is insufficient, as their continuous-time safety conditions do not directly guarantee constraint satisfaction over large discrete sampling intervals (Section 4.2). ASAP addresses this limitation through a discrete-time affine barrier constraint (Eq. 12), which guarantees safety at the next generated sample under the conditions of Theorem 4.1. This enables efficient arbitrary-step safe generation with fewer safety corrections, including a single MeanFlow prediction and one correction for one-step generation.

To further validate this distinction, we conduct an additional Maze2d ablation with few-step adaptations of existing safety-guaranteed methods. The results are summarized in the table below (Table 9 in the revised main paper), with complete results across all benchmarks provided in Appendix C.8. With minimal sampling or correction steps ((1) in table below), SAD-F, SFM, and S-Diff. exhibit safety violations of 0.570.57, 0.050.05, and 0.710.71, respectively. With their default sampling or correction settings ((D) in table below), as specified in the original papers and released code, all three achieve zero violations but require 35.8435.84 s, 15.5015.50 s, and 17.0817.08 s, respectively. Notably, even when SFM uses a single MeanFlow prediction, it still requires multiple subsequent safety corrections to achieve zero violations in our evaluation. In contrast, ASAP achieves zero violations with one MeanFlow prediction and one correction in 0.060.06 s. These results demonstrate the importance of ASAP's arbitrary-step safety formulation rather than simply reducing the number of model evaluations.

Comparison of safety corrections under one-step sampling (1) and default smpling steps (D) on Maze2d

Methods-vio(↓\downarrow)time(↓\downarrow)score(↑\uparrow)
SAD-F(1)0.57±0.360.57\pm0.360.10±0.010.10\pm0.011.61±0.071.61\pm0.07
SAD-F(D)0.00±0.000.00\pm0.0035.84±2.5135.84\pm2.511.44±0.281.44\pm0.28
SFM(1)0.05±0.080.05\pm0.080.11±0.000.11\pm0.000.76±0.660.76\pm0.66
SFM(D)0.00±0.000.00\pm0.0015.50±0.7515.50\pm0.751.30±0.671.30\pm0.67
S-Diff.(1)0.71±0.200.71\pm0.200.86±0.180.86\pm0.180.80±0.570.80\pm0.57
S-Diff.(D)0.00±0.000.00\pm0.0017.08±2.2917.08\pm2.291.46±0.231.46\pm0.23
ASAP (1)0.00±0.000.00\pm0.000.06±0.010.06\pm0.011.46±0.471.46\pm0.47

Regarding computational overhead, Eq. (13) admits a closed-form solution for independent state-wise affine constraints, as derived in Eq. (26). The original results in Tables 2–3 use a general-purpose QP solver for greater flexibility in handling multiple affine safety constraints, where deriving closed-form solutions is generally more challenging. Nevertheless, we additionally evaluated the closed-form implementation (ASAP (CF)) across all benchmarks. As summarized below for Maze2d, ASAP (CF) reduces planning time from 0.06 to 0.01 s while maintaining identical safety and task performance. Complete results are provided in Appendix C.7 (Table 16) of the revised manuscript.

Ablation of closed-form implementation (Maze2d)

Methods-vio(↓\downarrow)time(↓\downarrow)score(↑\uparrow)
ASAP (CF)0.00±0.000.00\pm0.000.01±0.000.01\pm0.001.46±0.471.46\pm0.47
ASAP (QP)0.00±0.000.00\pm0.000.06±0.010.06\pm0.011.46±0.471.46\pm0.47
AuthorsOpenReview ↗

Response to Reviewer YWCW (3/5)

Q3:Role of intermediate relaxation and hˉ\bar h

Clarify the purpose of the intermediate relaxation schedule, how hˉ\bar h is determined across trajectory waypoints, and whether the tested values satisfy the theoretical condition.

We thank the reviewer for this question. We agree that terminal safety in Theorem 4.1 depends only on ρtN=0\rho_{t_N}=0. The intermediate schedule ρt\rho_t instead provides progressive safety guidance during multi-step generation. Specifically, hˉ\bar h controls the relaxation level throughout generation, starting from ρt0=−hˉ\rho_{t_0}=-\bar h and progressively tightening toward zero, as defined in eq. (10). This allows intermediate constraint violations while guiding trajectories toward the safe set. This can avoid substantial corrections and improve task performance, as discussed in the previous question and included in Table 7 in revised paper.

The condition on hˉ\bar h is applied across all trajectory waypoints, i.e., hˉ≥max⁡k{−h(τt0(k)),0}\bar h\geq\max_k\{-h(\boldsymbol{\tau}_{t_0}(k)),0\}. In Maze2d, the circular obstacles have radius 11, so the maximum possible violation magnitude is 11, attained at the obstacle center. We further verified that the maximum observed initial violation is 0.880.88. Therefore, all tested values hˉ∈1,2,3\bar h\in\\{1,2,3\\} satisfy the theoretical condition. As shown in Table 5, ASAP maintains zero safety violations and similar task scores across these values. We have clarified this in the revised manuscript.

Q4:KUKA constraint settings and evaluation protocol.

Clarify the constraint levels used across Table 2, Table 10, and Fig. 3b, including the number of seeds/episodes and how the mean and standard deviation are computed.

Table 2 uses the 10%10\% constraint level for KUKA. Fig. 3(b) evaluates increasingly strict constraints (10%10\%, 20%20\%, and 30%30\%), while Table 10 compares additional baselines at the 30%30\% level. As reported in Appendix Table 18, KUKA is evaluated over 20 episodes, with mean and standard deviation computed across these episodes. We have clarified the constraint level in Section 5.1.

W2-2+Q5:Maze2d task-score sensitivity to safety constraints.

Clarify why MF and ASAP achieve identical task scores, why scores remain unchanged across (\bar h), and whether obstacle constraints affect the task evaluation.

Identical Maze2d scores for MF and ASAP. Following the D4RL benchmark [1], the Maze2d reward evaluates goal-reaching performance, while collision avoidance is measured separately by safety violation (s-vio). The introduced obstacles restrict feasible trajectories without blocking the goal. Thus, ASAP maintains nearly identical task scores to MF (with minor differences beyond the reported precision) while eliminating safety violations. This demonstrates that ASAP can enforce unseen safety constraints without compromising goal-reaching performance.

Sensitivity to hˉ\bar h (Table 5). The parameter hˉ\bar h controls the intermediate barrier relaxation schedule. We empirically observe that, when the theoretical assumption is satisfied, ASAP is relatively insensitive to hˉ\bar h within the tested range, generating nearly identical trajectories and task scores while maintaining safety.

Q6:Real-world collision evaluation and trajectory tracking.

Clarify whether MF's safety violations resulted in physical collisions, how the normalized collision margin is measured and the robot tracks trajectories.

For hardware safety, we remove the physical juicer during potentially unsafe executions to prevent robot damage, allowing MF to complete the task despite generating trajectories that would collide with the obstacle. Since the obstacle geometry is known, we retain its SDF to compute the normalized safety violation (s-vio), even when the physical obstacle is removed. Thus, MF achieves a task score of 1.001.00 despite its safety violation of 0.350.35. We have added the mathematical details of this metric in Section 5.3 of the revised manuscript. The generated trajectories are tracked using a low-level impedance controller, as also clarified in Section 5.3.

Q7:Multiple-obstacle SDFs and differentiability.

Clarify how ASAP handles overlapping obstacles and non-unique nearest-boundary points, and whether these cases occur in the experiments.

As defined in Appendix A, Eqs. (16)–(18), ASAP uses the union SDF for unsafe regions, which accommodates overlapping obstacles. Eq. (19) provides a simplified representation min⁡jhj\min_j h_j for non-overlapping obstacles, as used in our experiments. We have clarified this distinction in the revised manuscript. For non-unique nearest boundaries, a valid nearest-boundary direction can be selected for safety correction, as discussed in Remark 4.1. In our experiments, the introduced obstacles do not overlap, and no non-unique nearest-boundary points were observed.

[1] Fu, Justin, et al. "D4rl: Datasets for deep data-driven reinforcement learning." arXiv preprint arXiv:2004.07219, 2020.

Reviewer_nATVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ASAP, a generative planning framework that enables fast and safe trajectory generation with an arbitrary number of sampling steps. At each generative step, an affine barrier constraint based on a signed distance function (SDF) and a progressively tightened threshold is enforced through a quadratic program. ASAP also provides formal safety guarantees while accommodating previously unseen constraints without retraining. Experiments across a wide variety of tasks demonstrate substantial improvements in planning efficiency while maintaining safety and competitive task performance.

Critical strengths

(1) High planning efficiency with flexible safety enforcement. By combining fast MeanFlow generation with lightweight safety corrections, ASAP achieves substantial improvements in planning efficiency compared to existing safety-guaranteed methods. The ability to enforce previously-unseen safety constraints without retraining also strengthens its practical applicability.

(2) Intuitive and theoretically-sound safety mechanism. The proposed SDF-based correction is mathematically straightforward and provides formal safety guarantees for generated trajectory states, regardless of the number of sampling steps or the accuracy of the learned generator. Even though the underlying geometric principles are not entirely novel, their integration into an efficient generative AI planning framework is valuable.

(3) Extensive experimental validation. The paper evaluates ASAP across a variety of tasks in both simulation and hardware (a real-world Franka robot handover task). They demonstrate consistently strong results in many different performance metrics, e.g., safety, planning efficiency, overall task performance.

Critical weaknesses

(1) Limited novelty and justification of the proposed safety mechanism. Previous methods such as SafeFlowMatcher already separate nominal trajectory generation from subsequent safety correction. ASAP introduces a more efficient SDF-based correction that guarantees terminal safety with a single QP solve. However, the same guarantee could be achieved by applying this correction only to the final generated trajectory, without correcting intermediate generative steps. The benefits of these intermediate corrections are not theoretically established or isolated through experimental ablations.

(2) Insufficient ablation studies and robustness evaluation. Given the above concern (1), I worry the reported efficiency improvements largely result from the fast MeanFlow backbone than the proposed safety mechanism. Although Table 4 compares against direct post-optimization, simpler projection-based alternatives (like a final-only SDF projection scheme) are not investigated. Perhaps another alternative could be an analytical SDF projection, which raises further questions about the the necessity of a general QP solver. The hyperparameter study also only considers one parameter in a limited setting. There is little to no sensitivity analyses examining SDF approximation errors, numerical inaccuracies, or substantial trajectory corrections. The lower task performance on Adroit compared to some safety-guaranteed baselines also raises questions about the overall safety-performance tradeoff.

(3) Unclear applicability of the theory assumptions. Theorem 4.1 establishes safety at the sampled states of a generated trajectory under assumptions like exact differentiable SDFs and exact safety corrections. While these simplifying assumptions are reasonable to make the theory tractable, their relationship to the practical experiments is not sufficiently discussed. Furthermore, the formal guarantees do not directly address safety between consecutive trajectory states or consistency with the robot's physical dynamics.

Questions for authors

(1) Necessity of intermediate safety corrections. Given that terminal safety could be guaranteed by a final-only SDF projection under Theorem 4.1, could the authors compare ASAP against this simpler alternative? Comparing safety, task performance, correction magnitude, and runtime would help especially clarify the benefits of intermediate corrections. For independent statewise constraints, could the authors also explain the necessity of the QP solver instead of direct analytical SDF projection?

(2) Breakdown of ASAP versus MeanFlow contributions. How much of ASAP's reported speedup comes from the MeanFlow backbone versus the actual proposed safety correction mechanism? Likewise, does the lower task performance on Adroit, compared to some safety-guaranteed baselines, result primarily from MeanFlow or the safety correction mechanism?

(3) Applicability of theoretical assumptions. Are the assumptions made in the theoretical analyses intended primarily as sufficient conditions for theoretical analysis? If so, could the authors clarify which assumptions are satisfied / relaxed in each experimental setting, particularly in the real-world robot experiment? Also, how robust is the proposed method to SDF approximation errors, inaccurate gradients, and numerical solver tolerances?

(4) Scope of the formal safety guarantees. Further clarifying the distinction between the guarantees of Theorem 4.1 and the empirically observed safety performances would be helpful. For instance, does the safety certification cover only the discrete generated trajectory states, or does it also extend to the interpolated and physically executed trajectories? How are dynamic feasibility, action consistency, and tracking errors handled after safety correction?

(5) Hyperparameter robustness. Table 5 examines only the initial relaxation parameter in a limited setting. How sensitive is ASAP to other design choices (e.g., relaxation schedule, scenarios requiring substantial trajectory corrections)?

(6) Minor comments:

  • eq(6) defines a MeanFlow identity using a velocity conditioned on the interval endpoint, whereas eq(7) appears to evaluate the average velocity at its starting state. Are different time or conditioning conventions being used here?
  • eq(4) is missing the trajectory-state index in the control input.
  • Section 3.2: "by minimally deviates from the expert trajectory distribution" --> "by minimally deviating from the expert trajectory distribution"
7 comments
AuthorsOpenReview ↗

Response to Reviewer nATV (1/4)

We thank the reviewer for recognizing ASAP's efficiency, theoretical safety guarantees, and extensive experimental validation. We also appreciate the constructive questions, which helped us clarify our contributions and strengthen the manuscript.

W1+Q1: Necessity of ASAP's safety correction and QP solver.

Clarify the advantages of intermediate safety guidance over final-only SDF projection to justify ASAP's novelty, and explain the use of a general-purpose QP solver instead of a closed-form correction.

The main contribution of ASAP is a unified safety framework supporting both efficient one-step and multi-step generation. As shown in Fig. 3(a), complex tasks such as KUKA benefit from multiple sampling steps (N>1N>1), motivating the need for flexible safety guidance. While we agree that terminal safety can also be achieved through final-only SDF projection, ASAP incorporates a time-varying relaxation ρt\rho_t, motivated by prescribed-time safety control [1], that progressively guides generation toward safety without requiring full constraint satisfaction at intermediate steps. Unlike final-only projection, these relaxed corrections influence subsequent MeanFlow predictions, potentially avoiding substantial corrections at the final step.

To validate the benefit of intermediate guidance, we conducted additional ablations comparing ASAP with (i) MeanFlow followed by final-only SDF projection (Final-Only) and (ii) MeanFlow with full SDF projection at every step (Full). As summarized below, all methods satisfy the safety constraints, while ASAP achieves the highest task scores on KUKA ((N=4N=4): 1.751.75 vs. 1.151.15/1.351.35) for Final-only/Full) and real-world handover ((N=2N=2): 1.001.00 vs. 0.800.80/0.700.70). These results demonstrate the benefit of relaxed intermediate guidance over the alternative projection strategies. Complete results are reported in Table 7 and Appendix C.6 of the revised manuscript, including the equivalent safety and task performance of one-step methods.

Ablation of score result for SDF projection strategies

BenchmarkFinal-OnlyFullASAP
Kuka(N=4)1.15±0.791.15\pm0.791.35±0.961.35\pm0.961.75±0.831.75\pm0.83
Real robot(N=2)0.80±0.400.80\pm0.400.70±0.460.70\pm0.461.00±0.011.00\pm0.01

For independent state-wise constraints, we agree that the minimum-norm QP admits an analytical SDF correction. Our original implementation uses a general-purpose QP solver for its flexibility in handling multiple affine safety constraints, where closed-form solutions are generally more challenging. Nevertheless, we additionally evaluated the closed-form implementation (ASAP (CF)) across all benchmarks. As summarized below for Maze2d, ASAP (CF) reduces planning time from 0.060.06 to 0.010.01 s while maintaining identical safety and task performance. Complete results are provided in Appendix C.7 (Table 16) of the revised manuscript.

Ablation of closed-form implementation (Maze2d)

Methods-vio(↓\downarrow)time(↓\downarrow)score(↑\uparrow)
ASAP (CF)0.00±0.000.00\pm0.000.01±0.000.01\pm0.001.46±0.471.46\pm0.47
ASAP (QP)0.00±0.000.00\pm0.000.06±0.010.06\pm0.011.46±0.471.46\pm0.47

[1] Song, Yongduan, et al. "Time-varying feedback for regulation of normal-form nonlinear systems in prescribed finite time." Automatica 83: 243-251, 2017.

AuthorsOpenReview ↗

Response to Reviewer nATV (2/4)

W2-1+Q2: Speedup attribution of MeanFlow vs. ASAP and Adroit performance.

Clarify how ASAP's safety correction contributes to the reported efficiency beyond MeanFlow's fast generation, and whether the lower task performance on Adroit results from the MeanFlow backbone or the safety correction mechanism.

We thank the reviewer for this important question. We agree that MeanFlow's few-step sampling contributes substantially to ASAP's efficiency. However, ASAP's key contribution is enabling safe few-step generation with formal guarantees without introducing expensive additional overhead. While MeanFlow provides fast generation, it does not guarantee safety. Existing safe planners, including SAD-Flower (SAD-F), SafeFlowMatcher (SFM), and SafeDiffuser (S-Diff.), generally rely on repeated safety corrections to achieve safety, introducing substantial computational overhead. Simply reducing these corrections to match MeanFlow's few-step generation is insufficient, as their continuous-time safety conditions do not directly guarantee constraint satisfaction over large discrete sampling intervals (Section 4.2). ASAP addresses this limitation through a discrete-time affine barrier constraint (Eq. 12), which guarantees safety at the next generated sample under the conditions of Theorem 4.1. This enables efficient arbitrary-step safe generation with fewer safety corrections, including a single MeanFlow prediction and one correction for one-step generation.

To further validate this distinction, we conduct an additional Maze2d ablation with few-step adaptations of existing safety-guaranteed methods. The results are summarized in the table below (Table 9 in the revised main paper), with complete results across all benchmarks provided in Appendix C.8. With minimal sampling or correction steps ((1) in table below), SAD-F, SFM, and S-Diff. exhibit safety violations of 0.570.57, 0.050.05, and 0.710.71, respectively. With their default sampling or correction settings ((D) in table below), as specified in the original papers and released code, all three achieve zero violations but require 35.8435.84 s, 15.5015.50 s, and 17.0817.08 s, respectively. Notably, even when SFM uses a single MeanFlow prediction, it still requires multiple subsequent safety corrections to achieve zero violations in our evaluation. In contrast, ASAP achieves zero violations with one MeanFlow prediction and one correction in 0.060.06 s. These results demonstrate the importance of ASAP's arbitrary-step safety formulation rather than simply reducing the number of model evaluations.

Comparison of safety corrections under one-step sampling (1) and default smpling steps (D) on Maze2d

Methods-vio(↓\downarrow)time(↓\downarrow)score(↑\uparrow)
SAD-F(1)0.57±0.360.57\pm0.360.10±0.010.10\pm0.011.61±0.071.61\pm0.07
SAD-F(D)0.00±0.000.00\pm0.0035.84±2.5135.84\pm2.511.44±0.281.44\pm0.28
SFM(1)0.05±0.080.05\pm0.080.11±0.000.11\pm0.000.76±0.660.76\pm0.66
SFM(D)0.00±0.000.00\pm0.0015.50±0.7515.50\pm0.751.30±0.671.30\pm0.67
S-Diff.(1)0.71±0.200.71\pm0.200.86±0.180.86\pm0.180.80±0.570.80\pm0.57
S-Diff.(D)0.00±0.000.00\pm0.0017.08±2.2917.08\pm2.291.46±0.231.46\pm0.23
ASAP (1)0.00±0.000.00\pm0.000.06±0.010.06\pm0.011.46±0.471.46\pm0.47

For Adroit, we investigated the original performance gap and found that the competing baselines employ reward-based candidate selection, whereas neither MeanFlow (MF) nor ASAP included this procedure in our original experiments. Thus, the original performance gap cannot be attributed solely to ASAP's safety correction, since uncorrected MeanFlow also achieved a lower score than the competing safe baseline. We therefore conducted additional experiments incorporating the same reward-based selection procedure, improving ASAP's task score to 1.101.10 as the table below. We have clarified the evaluation protocol and updated the Adroit results in the revised manuscript (Table 6.).

Updated Adroit result

MetricMFS-Diff.ASAP
s-vio(↓\downarrow)0.35±0.070.35\pm0.070.00±0.000.00\pm0.000.00±0.000.00\pm0.00
time(↓\downarrow)0.02±0.000.02\pm0.000.59±0.010.59\pm0.010.15±0.000.15\pm0.00
score(↑\uparrow)1.07±0.121.07\pm0.121.07±0.081.07\pm0.081.10±0.061.10\pm0.06
AuthorsOpenReview ↗

Response to Reviewer nATV (3/4)

W3-1+Q3: Applicability of theoretical assumptions and robustness.

Clarify how the theoretical assumptions apply to practical experiments, particularly real-world robotics, and evaluate robustness to SDF approximation errors, inaccurate gradients, and solver tolerances.

We thank the reviewer for these questions and address them as follows.

(1) Applicability of theoretical assumptions. The assumptions in Theorem 4.1 provide sufficient conditions for theoretical safety guarantees. In our experiments, safety constraints are represented using geometric SDFs. In particular, for the real-world robot, we use enlarged ellipsoid approximations to conservatively enclose obstacle geometry, providing additional safety clearance, as commonly adopted in robotics [1–3]. Regarding differentiability, geometric SDFs are differentiable almost everywhere, with nondifferentiable points forming a set of measure zero, and no such points were encountered during planning in our experiments. We have added further details on the constraint formulation in Section 5.3 of the revised manuscript.

(2) Robustness to practical inaccuracies. To account for practical inaccuracies arising from SDF approximation, inaccurate gradients, and numerical solver tolerances, we extend ASAP by assuming their combined effect on the barrier constraint is bounded by eˉ\bar e. We introduce a linearly increasing robustness margin rti=tieˉr_{t_i}=t_i\bar e, extending Eq. (12) to

\hat{\boldsymbol{g}}(\boldsymbol{\hat{\tau}}_{t_i})^\top \left(\boldsymbol{u}_{t_i}\Delta t\right) \geq \rho_{t_{i+1}} + r_{t_{i+1}}$$ where $\hat h$ and $\hat{\boldsymbol{g}}$ denote the estimated SDF and gradient. Assuming the total error between the prescribed affine barrier threshold and the true barrier at the corrected state, including numerical constraint violations, is bounded by $\bar e$, we obtain $h(\boldsymbol{\tau}_{t_{i+1}}) \geq \rho_{t_{i+1}} + r_{t_{i+1}} - \bar{e}.$ At the terminal step, $r_1=\bar e$ and $\rho_1=0$, yielding $h(\boldsymbol{\tau}_1)\geq0$. Thus, ASAP preserves terminal safety for arbitrary sampling steps under bounded integrated errors, provided the robust constraint is feasible. We provide the theorem in **Section 4.2 (Proposition 4.1)** and its proof in the revised appendix B.4. **(3) Empirical robustness evaluation.** We further evaluate the robust formulation on Maze2d under Gaussian noise with standard deviations $\sigma\in\\{0.1,0.4,0.8\\}$, using a robustness parameter $\bar e=1$. As shown in Table below, ASAP maintains zero safety violations across all tested noise levels, with similar task scores ($1.41–1.42$) and planning times ($0.05–0.06$ s). These results demonstrate the empirical robustness of ASAP under the tested perturbations with minimal computational overhead. This ablation is included in the revised paper (Table 8.). **Robust ASAP in Maze2d** | $\sigma$ | s-vio($\downarrow$) | time($\downarrow$) | score($\uparrow$) | | -------- | -------- | -------- | -------- | | $0.1$ | $0.00\pm0.00$ | $0.05\pm0.01$ | $1.42\pm0.46$ | | $0.4$ | $0.00\pm0.00$ | $0.05\pm0.01$ | $1.41\pm0.46$ | | $0.8$ | $0.00\pm0.00$ | $0.06\pm0.01$ | $1.41\pm0.47$ | [1] Liu, Baolin, et al. "Collision-free motion generation based on stochastic optimization and composite signed distance field networks of articulated robot." IEEE Robotics and Automation Letters 8.11: 7082-7089, 2023. [2] Li, Yiming, et al. "Representing robot geometry as distance fields: Applications to whole-body manipulation." ICRA, 2024. [3] Hu, Songqiao, et al. "Vlsa: Vision-language-action models with plug-and-play safety constraint layer." IROS, 2026.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Author Response and Manuscript Revisions

Dear AC and reviewers,

We sincerely thank all reviewers for their thoughtful comments. We appreciate the reviewers' recognition of ASAP's contributions and have carefully addressed their concerns through theoretical clarifications, additional experiments, and manuscript revisions. We summarize the key points below.

1. Strengths Highlighted by Reviewers

We appreciate the reviewers' positive assessment of:

  • Efficient safety-guaranteed planning. ASAP enables one-step and few-step trajectory generation with formal safety guarantees, including for previously unseen constraints without retraining.
  • Clear theoretical formulation. The discrete-time barrier formulation provides rigorous safety guarantees for arbitrary sampling steps under the stated assumptions.
  • Extensive experimental validation. ASAP demonstrates zero planning-level safety violations, competitive task performance, and substantial efficiency improvements across simulation benchmarks and real-world robotic handover.

2. Key Clarifications

  • Benefits of ASAP's intermediate safety projection strategy. ASAP provides a unified safety framework for both one-step and multi-step generation. While one-step ASAP is equivalent to final-only SDF projection under exact-SDF assumptions, multi-step generation benefits from ASAP's time-varying safety relaxation. Unlike final-only projection, which corrects trajectories only at the end, ASAP's intermediate corrections influence subsequent MeanFlow predictions, potentially avoiding substantial final corrections. Unlike full projection, ASAP allows intermediate constraint violations and progressively guides trajectories toward safety. Our additional ablations demonstrate the benefit of this flexible guidance, achieving higher task scores on KUKA (1.751.75 vs. 1.151.15/1.351.35) and real-world handover (1.001.00 vs. 0.800.80/0.700.70) compared with final-only/full projection, with all methods maintaining safety.
  • Efficiency beyond the MeanFlow backbone. MeanFlow enables fast generation but does not guarantee safety. Existing safe planners, i.e., SafeDiffuser, SafeFlowMatcher, and SAD-Flower, generally require repeated corrections, and their continuous-time safety guarantees do not directly extend to large discrete sampling intervals. ASAP addresses this limitation through a discrete-time affine barrier constraint, enabling formally safe generation with arbitrary sampling steps and fewer corrections. Our new ablations confirm that few-step adaptations of existing safe planners exhibit safety violations, whereas ASAP achieves zero violations with a single MeanFlow prediction and one correction.

3. Revisions in the Manuscript

  • Projection strategy ablations (Table 7, Appendix C.6): Comparisons against final-only and full SDF projection across all benchmarks.
  • Few-step baseline comparisons (Table 9, Appendix C.8): Evaluation of SAD-Flower, SafeFlowMatcher, and SafeDiffuser under few-step and default multi-step settings.
  • Closed-form implementation (Appendix C.7): Analytical safety corrections that further reduce planning time without changing the safety formulation.
  • Robustness analysis (Proposition 4.1, Table 8): A bounded-error safety formulation and experiments evaluating robustness to Gaussian noise.
  • Expanded experimental validation: Additional real-world baselines (Appendix C.3), revised Adroit comparisons with aligned candidate selection (Table 6).
  • Clarified theoretical scope and problem formulation: Clarification of SDF assumptions (Appendix A), and refinement of the objective with competitive task performance (Section 3.2).

We believe these revisions provide stronger theoretical clarity and experimental evidence for ASAP's effectiveness in efficient, safety-guaranteed generative planning.

AuthorsOpenReview ↗

Summary of Responses For 2nd Round Discussion

Dear AC and reviewers,

Following the second-round discussion, we would like to further emphasize that ASAP goes beyond standard final-only SDF projection and prediction–correction frameworks through two key advances.

(1) Beyond final-only projection. Final-only SDF projection is a special case of ASAP under one-step generation (N=1N=1), under the assumptions of Theorem 4.1. However, terminal safety alone does not establish equivalent planning performance. In tasks where few-step generation yields better planning performance, ASAP's relaxed intermediate corrections guide subsequent MeanFlow predictions, potentially avoiding substantial final corrections. Our additional ablations show that ASAP achieves 5252% higher task scores on KUKA (N=4N=4) and 2525% on real-world handover (N=2N=2) than final-only projection, with both maintaining terminal safety. These results demonstrate that ASAP's intermediate safety guidance achieves superior task performance beyond terminal projection alone, and is reported in Table 7 of the revised paper.

(2) Beyond existing safety correction frameworks. ASAP's discrete-time affine barrier formulation provides formal safety guarantees for arbitrary sampling steps, whereas existing continuous-time safety conditions do not directly guarantee safety over large discrete intervals. Reducing existing methods to one-step generation leads to safety violations in our experiments, while increasing correction steps to achieve safety introduces substantial computational overhead. Specifically, SafeFlowMatcher and SafeDiffuser require 15.5015.50 s and 17.0817.08 s, respectively, under their default settings to achieve zero violations on Maze2d. In contrast, ASAP achieves zero violations in only 0.06 s with one MeanFlow prediction and one correction. These comparisons are reported in Table 9 of the revised manuscript.

Together, these results highlight ASAP's methodological contributions in arbitrary-step safety formulation and intermediate guidance, as well as its empirical advances in planning efficiency and task performance. We believe these distinctions clarify the contribution of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.