acceptodds
Under review as a conference paper at ICLR 2027

ANCHOR: Safe Diffusion Planning via Analytic Approximation of Doob’s -function

Abstract

Diffusion models provide expressive, multimodal priors for offline trajectory planning, but enforcing hard safety constraints without distorting the learned trajectory distribution remains challenging. We formulate safe planning as conditional generation from the distribution of trajectories given collision avoidance, whose exact reverse process is characterized by a Doob \(h\)-transform. Since the corresponding safety posterior probability is generally intractable, we introduce ANCHOR, a training-free analytic approximation to the Doob guidance. Our method approximates the posterior using Tweedie moments and exploits signed-distance geometry for efficient computation. Under explicit regularity assumptions on the data distribution and obstacle geometry, we prove that the resulting guided diffusion is almost surely safe. On synthetic multimodal distributions, our method recovers the safety-conditioned distribution substantially more accurately than safe generative-planning baselines while avoiding boundary concentration. On multimodal navigation, locomotion, and manipulation benchmarks it achieves results that outperform or remain competitive with state-of-the-art safe diffusion and flow-matching planners.

Primary area:
generative models
Keywords:
Diffusion Models, Probabilistic Methods, Safe Planning, Robotics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9u41OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper treats safe trajectory planning with a pretrained diffusion planner as sampling from p0(x0∣x0∈S)p_0(x_0 \mid x_0 \in S). The exact reverse process for this target is the Doob hh-transform, with guidance ∇log⁡ht\nabla \log h_t and ht(x)=P(X0∈S∣Xt=x)h_t(x) = P(X_0 \in S \mid X_t = x). ANCHOR approximates hth_t in closed form in two steps. (i) It approximates p(x0∣xt)p(x_0 \mid x_t) by a Gaussian with the Tweedie mean and, in practice, the isotropic covariance rt2Ir_t^2 I. (ii) It linearizes the obstacle's signed-distance function (or a normalized level-set surrogate) at the Tweedie mean. This reduces the safe mass to a 1-D Gaussian tail Φ(sC(x^0)/σ)\Phi(s_C(\hat x_0)/\sigma). The guidance is ∑j∇log⁡Φ(sC(x^0(j))/rt)\sum_j \nabla \log \Phi(s_C(\hat x_0^{(j)})/r_t) over waypoints, and it is active only for t≤tmax⁡t \le t_{\max}. No retraining is needed. Theorem 2 proves almost-sure safety for half-space obstacles. It assumes the data distribution is a Gaussian with block-diagonal covariance times a tilt with bounded gradient, the exact score, continuous time, and tempered guidance γft\gamma f_t with γ>1\gamma > 1.

Validation. (a) 1-D and 2-D Gaussian mixtures and a three-corridor synthetic trajectory task, where the safe conditional is known (W1/SW1, safety, boundary excess). (b) Two new Multimodal Maze2D datasets (Medium: 3 modes; Hard: 6 modes) with test-time obstacles that block modes. (c) Hopper and Walker2d with root-height ceilings, and Kuka with 90% joint limits. Baselines: Diffuser, classifier guidance, Truncate, SafeDiffuser (RoS/ReS/TVS), Flow Matching, SafeFlowMatcher, and SAD-Flower. ANCHOR gets the lowest distributional error at 100% safety on the synthetic tasks. It gets much higher planned safety on Multimodal Maze (Medium 94.3% vs. at most 43.8%; Hard 45.3% vs. at most 5.7%). It reaches 100% planned safety on locomotion and Kuka with similar task scores. Table 3 also shows that in the original benchmark settings few or no training trajectories satisfy the test constraints.

Critical strengths
  1. A principled formulation with a cheap closed-form guidance, and an evaluation that measures the right thing. Treating safety as conditioning separates two questions: "is the plan safe?" and "is the plan still consistent with the data distribution?". The paper measures distributional fidelity (W1/SW1, boundary excess) in addition to safety rate, which most prior safe-diffusion work does not do. The results support the claim. On the 1-D mixture, W1 is 0.059 vs. 0.629 for the best 100%-safe baseline (SafeDiffuser-RoS). On the three-corridor task, safety is 100% vs. at most 52.3%. On Multimodal Maze Medium it is 94.3% vs. at most 43.8%, and the plans visibly reroute through the free corridor instead of hugging the obstacle (Fig. 11, trap rates in Table 7). This is the main reason I lean toward acceptance.

  2. A useful benchmark analysis, plus non-trivial theory. Table 3 shows that in the original Maze2D setting and the original locomotion ceilings, 0% (Maze2D, Hopper) or about 5% (Walker2d) of training windows satisfy the test constraints. So those benchmarks test extrapolation, not conditional sampling. This clarifies what "safe diffusion planning" benchmarks actually measure, and it will be useful to the community. The barrier/supermartingale proof of Theorem 2 is also a non-trivial result: a Gaussian-approximated hh-transform stays safe under a bounded tilt of the data distribution.

Critical weaknesses
  1. The safety theorem does not cover the method that is evaluated, and gains on the standard robotics tasks are small.

    • Theorem 2 needs half-space obstacles, tilted-Gaussian data, the exact score, continuous time, γ>1\gamma > 1, and it assumes that the guided SDE has a solution.
    • All experiments use γ=1\gamma = 1, and the authors note that the proof gives nothing for γ=1\gamma = 1. They also use curved obstacles, a learned score with 256 discrete steps, and, on Maze, a norm cap of 2 on ftf_t. A bounded guidance removes the blow-up that the barrier argument relies on.
    • Empirically, planned safety on Maze Hard is 22.9-45.3%, and execution safety on Hopper/Walker2d is 39%/60%.
    • With γ>1\gamma > 1 the guidance also departs from the Doob transform, so the theorem is about safety only, not about recovering p0(x0∣S)p_0(x_0 \mid S). The abstract's "we prove ... almost surely safe" should be scoped to these conditions.
    • On the standard robotics tasks, ANCHOR is close to simple Truncate. Hopper: 0.437 vs. 0.422 (SD about 0.09-0.10), both 100% planned and 39% execution safety. Walker2d: 0.226 vs. 0.209, and Truncate has higher execution safety (73% vs. 60%). Kuka: Truncate has higher reward (0.710 vs. 0.680). So the clear benefit shows mainly on the synthetic tasks and the authors' new Maze datasets.
  2. A key baseline is missing, and the tuning protocol is unclear.

    • Rejection sampling from the unguided planner samples p0(x0∣x0∈S)p_0(x_0 \mid x_0 \in S) exactly. It is the natural reference for the paper's own objective. Base acceptance rates are about 37% (Maze Medium), 51-58% (locomotion, planned), and 48% (Kuka), so a few draws are enough. Only Maze Hard (about 4.7%) is costly.
    • Training-free guidance baselines that use the same SDF on the Tweedie mean (e.g., DPS / universal guidance with a hinge or log-barrier loss) and projection methods (Christopher et al., 2024; Römer et al., 2025) are also missing.
    • Runtime is not reported, although the conclusion claims competitive computational cost.
    • ANCHOR's configuration (tmax⁡t_{\max}, time weight, gradient type, norm cap) changes per benchmark (Table 4). The sweeps in Tables 9 and 12-14 appear to use the same evaluation episodes. On Maze, two configurations are reported and the best of each is bolded (constant weight for reward 77.73, inverse-std weight for safety 45.3%). The baselines were modified (SAD-Flower moved to a U-Net and evaluated on the first sample instead of best of 32; SafeFlowMatcher with Tp=16T_p = 16) but do not seem to get similar tuning.
Questions for authors

Clarifications (could change my assessment):

  1. Were ANCHOR's hyperparameters (tmax⁡t_{\max}, mtm_t, gradient type, norm cap) chosen on held-out episodes or seeds, or on the reported evaluation set? Can you report one fixed configuration for all tasks, or a validation-based selection?
  2. What is the wall-clock planning time compared with Diffuser, SafeDiffuser, and SafeFlowMatcher, for both the through-model and stopped-model gradients?
  3. Does any safety argument survive the norm cap used on Maze? Can you show, even only empirically, how safety changes with γ\gamma and with the number of sampling steps?
  4. Theorem 2 assumes that the guided SDE has a solution with continuous paths. Is this guaranteed for the proposed drift, which grows like ∣d∣/rt2|d|/r_t^2 inside the obstacle?
  5. Which noise schedule and time normalization do the experiments use? tmax⁡∈[0,1]t_{\max} \in [0, 1] does not match the OU process of Eq. (1) with large TT.

Experiments that would materially strengthen the paper:

  1. A rejection-sampling baseline (unguided Diffuser plus a safety check, KK draws) at matched compute on all benchmarks, including W1/SW1 on the synthetic tasks.
  2. Results in the original constraint settings (reference Maze2D, Hopper 1.5 m, Walker2d 1.3 m), where the safe conditional has no data support.
  3. A training-free guidance baseline that uses the same SDF on the Tweedie mean (e.g., a DPS-style gradient of a hinge or log-barrier loss).
6 comments

Response to Reviewer 9u41 (1/5)

We thank the reviewer for recognizing the principled formulation, scope, and theoretical results of our work. We also appreciate the constructive experiment suggestions and address the concerns and questions below; all new results will be added to the revised manuscript.

Three regimes. As the reviewer observes, none of the training trajectories satisfy the original benchmark constraints, so these settings test extrapolation rather than conditional sampling (Table 3 of the paper). We found it useful to organize our answers around three regimes, all of which appear in our benchmarks:

  • (i) supported: safe solutions are represented in the data, the target p0(x0∣x0∈S)p_0(\mathbf{x}_0\mid\mathbf{x}_0\in S) is well defined, and rejection sampling is a valid way to reach it (Gaussian mixtures, Sequential, Multimodal Maze Medium, tightened locomotion and Kuka constraints);
  • (ii) unsupported: no or almost no training trajectory is safe (original Maze2D, Hopper 1.5 m, Walker2d 1.3 m), the conditional is degenerate, and every method has to extrapolate;
  • (iii) mixed: the planner has to move to a supported safe mode and deviate locally from it (Multimodal Maze Hard, whose single open route is additionally half blocked).

Constraints in real deployments are typically of the third kind. ANCHOR handles all three with one guidance term. Rejection sampling is limited to regime (i), and local correction methods (CBFs, projection, truncation) to regime (ii).

1. On the scope of our theoretical safety guarantees (W1). We agree that the submitted Theorem 2 did not cover the method as evaluated. We have revised the proof (Appendix B.2), and it now closes most of the gaps the reviewer lists:

  • Guidance multiplier γ\gamma. The submitted proof required γ>1\gamma>1 for a pointwise drift condition; this was a theoretical artifact and not part of our method. We have revised our strategy and found a proof that works for γ=1\gamma = 1; details are given in item 2 of our response to Reviewer mDHV. It is correct that γ≠1\gamma\neq1 no longer corresponds to the Doob guidance for p0(x0∣x0∈S)p_0(\mathbf{x}_0\mid\mathbf{x}_0\in S), so we have removed it from the revised paper.
  • Obstacles and data. The theorem now additionally covers a bounded convex unsafe set per waypoint, next to the already existing half-space obstacle.
  • Existence of a solution (Q4). Theorem 2 no longer assumes that the guided SDE has a solution: Lemma 6 shows that it has a unique strong solution with continuous paths before the data end, which extends continuously to the data end, although the drift grows like ∣d∣/rt2\vert d\vert/r_t^2 inside the obstacle.
  • Implementation choices. The isotropic covariance rt2Ir_t^2 I and the stopped-model Jacobian are covered by the remark after Theorem 2.

Outside the theorem remain the standard approximations of any diffusion sampler (time discretization and a learned score in place of the exact one), several obstacles per waypoint combined through the minimum of their SDFs (the maze benchmarks, including walls), and non-convex unsafe sets (the complement of the joint-limit box on Kuka). The norm cap is addressed in item 5.

Regarding the number of sampling steps (Q3), the table below shows that planned safety on Maze Hard increases with the number of steps, from 22.9% at 256 steps to 52.1% at 512 steps, at the cost of lower reward and a smaller fraction of plans with 1 cm clearance; on the 2-D mixture, SW$_1$ decreases from 0.108 to 0.090 at 100% safety.

Sampling-step ablation of ANCHOR (constant weight); three seeds with 10,000 samples (GMM2D) and 192 plans (Multimodal Maze Hard).

StepsGMM2D SW$_1\downarrow$GMM2D Safe (%)Maze Hard, cap: RewardSafe (%)B-Safe (%)Maze Hard, no cap: RewardSafe (%)B-Safe (%)
1280.108 ± 0.004100.085.03 ± 56.3622.921.482.43 ± 57.3422.921.4
2560.100 ± 0.010100.077.73 ± 60.0522.919.877.34 ± 60.3823.421.4
5120.090 ± 0.005100.069.29 ± 62.7652.115.167.74 ± 63.8951.015.1

Response to Reviewer 9u41 (2/5)

2. On the empirical performance and the benchmark constraints (W1, Q7).

Standard robotics benchmarks. We agree that ANCHOR's gains over Truncate on Hopper, Walker2d, and Kuka are modest. These constraints are axis-aligned bounds (ceiling heights and joint limits), which Truncate enforces by clipping single coordinates, a correction that is close to optimal for box constraints. ANCHOR matches Truncate in planned safety and task score on these tasks and keeps larger margins (planned B-Safe at 1 cm: 100% vs. 43% on Hopper and 100% vs. 56% on Walker2d; at 10% on Kuka: 72% vs. 58%). Where a local correction is not sufficient, i.e. on Multimodal Maze, ANCHOR improves substantially over all baselines (94.3% vs. at most 43.8% on Medium, 45.3% vs. at most 5.7% on Hard). We acknowledge the remaining violations on Maze Hard and discuss them in the paper.

Execution safety on locomotion. Plans are executed with replanning under unmodeled dynamics, and execution safety is limited for all methods. The table below shows that it also conflates safety with falling: for every method, 56–100% of the execution-safe episodes terminate before 500 actions, i.e. the robot falls before it can touch the ceiling (≥98% for FlowMatcher and SafeFlowMatcher, whose safe episodes score at most 0.134).

Executed rollouts under the main-paper ceilings, 100 episodes per method with up to 500 actions. Safe: ceiling satisfied throughout the episode. Early: percentage of safe episodes that terminate before 500 actions. ANCHOR uses tmax⁡=0.3t_{\max}=0.3.

MethodHopper 1.6 m: Safe (%)Score (safe)Score (unsafe)Early (%)Walker2d 1.4 m: Safe (%)Score (safe)Score (unsafe)Early (%)
Diffuser360.3820.45572.2670.1370.35389.6
CG350.3740.46474.3710.1580.35085.9
Truncate390.3670.45759.0730.1520.36587.7
SafeDiffuser–RoS420.3720.45464.3700.1640.34288.6
Flow Matching670.3230.44891.01000.105–98.0
FlowMatcher970.1340.351100970.0510.25897.9
SafeFlowMatcher960.1340.349100950.0500.199100
ANCHOR (constant)410.3620.46373.2720.1630.35784.7
ANCHOR (inverse std)390.4060.45756.4600.1350.36290.0

Original constraint settings. As requested, we evaluated all methods under the original constraints (table below; full results in the revised appendix). Here, almost no training trajectory satisfies the constraint, so every method has to extrapolate and produce plans that are not in the data. On Maze2D, the gap between the lower obstacle and the wall is too narrow for the simulator's agent to pass on the main path, so no method can produce a safe plan along it. We therefore only report planned obstacle safety for a point agent, as in the original benchmark. On Hopper and Walker2d, the ceilings are lower than the height the robots reach while walking in the data (0% and 5.2% of training windows satisfy them). All safe planners reach 100% planned safety, but the robot cannot follow these plans, so execution safety is low. It is high only when the robot stops moving forward, e.g. ANCHOR with tmax⁡=1t_{\max}=1 on Hopper (100% at a score of 0.125).

Original constraints. Maze2D: 192 plans (three seeds), planned obstacle safety for a point agent, cumulative simulator reward. Hopper and Walker2d: 100 episodes, normalized score, planned/executed safety (%). ANCHOR uses inverse-standard-deviation weighting.

MethodMaze2D: RewardSafe (%)Hopper 1.5 m: ScoreSafe P/EWalker2d 1.3 m: ScoreSafe P/E
Diffuser413.56 ± 77.871.60.431 ± 0.0981/100.220 ± 0.1510/49
SafeDiffuser–RoS381.86 ± 117.2299.00.422 ± 0.101100/150.217 ± 0.149100/53
SafeFlowMatcher438.75 ± 2.59100.00.130 ± 0.061100/580.099 ± 0.051100/8
ANCHOR (tmax⁡=0.3t_{\max}=0.3)407.22 ± 97.01100.00.413 ± 0.105100/140.228 ± 0.156100/47
ANCHOR (tmax⁡=1t_{\max}=1)159.45 ± 96.46100.00.125 ± 0.020100/100−0.008 ± 0.004100/97

Response to Reviewer 9u41 (3/5)

3. On the additional baselines: rejection sampling, SDF guidance, and projection (W2, Q6/8).

Rejection sampling (RS). RS returns the first safe plan among up to K=16K=16 unguided draws. As K→∞K\to\infty it samples p0(x0∣x0∈S)p_0(\mathbf{x}_0\mid\mathbf{x}_0\in S) exactly as well as the model has learned it, so on the synthetic tasks it marks the error floor of the learned score (GMM1D/2D: W$_1$ 0.044, SW$_1$ 0.042; ANCHOR 0.059 and 0.071; next-best safe baseline ≥ 0.33). As a planner, its cost scales as 1/p1/p in the unguided acceptance rate pp, and its latency is random. It beats ANCHOR on Maze Medium (p≈0.37p\approx0.37; 99.0% vs. 94.3%). On Maze Hard (p≈0.05p\approx0.05), it falls behind (29.7% vs. 45.3%) at 5× the time. Under the original constraints (p≈0p\approx0), it fails, because almost no unguided plan is safe. ANCHOR still works there: its Gaussian posterior approximation puts mass everywhere, so the guidance remains well defined and acts as a local correction, and Theorem 2 holds however small P(S)P(S) is. The two approaches are complementary: ANCHOR raises the per-draw acceptance rate to 94% and 45%, which cuts the expected number of draws per safe plan from 2.7 to 1.1 (Medium) and from 21 to 2.2 (Hard). Locomotion and Kuka results are in the appendix.

SDF hinge guidance. We added the suggested DPS-style squared hinge

L(x^0)=∑jmax⁡{0,−s(x^0,j)}2L(\hat{\mathbf{x}}_0) = \sum_j \max\lbrace 0, -s(\hat{\mathbf{x}}_{0,j})\rbrace^2

(Chung et al., 2023; Bansal et al., 2023), using the same Tweedie mean and SDF as ANCHOR. Inside the obstacle (s≪−σts \ll -\sigma_t), ANCHOR's guidance ∇log⁡Φ(s/σt)\nabla\log\Phi(s/\sigma_t) approximately equals the hinge gradient with step λt=1/(2σt2)\lambda_t = 1/(2\sigma_t^2). It differs from the hinge in two ways, both set by the posterior variance rather than tuned:

  • Scale. The push is weak while the posterior is broad, which preserves mode proportions. It then grows like 1/σt21/\sigma_t^2, so its time integral diverges as t→0t\to0. This is the mechanism behind Theorem 2, while a fixed-step hinge stays bounded and has no such guarantee.
  • Margin. Waypoints just outside the boundary are still pushed, whereas the hinge stops at the boundary, which produces the boundary accumulation seen in Figs. 8–9.

On Multimodal Maze Hard, planned safety rises from 13.5% (hinge) to 22.9% (ANCHOR, constant weight) and 45.3% (ANCHOR, 1/σt1/\sigma_t weight). The hinge's higher overall reward includes unsafe plans. On Sequential, which reduces to mode selection, the two are tied. The appendix gives the derivation and the full ablation from hinge to ANCHOR.

From hinge to ANCHOR. Within each benchmark, all rows share the Tweedie mean, SDF, gradient path, tmax⁡t_{\max} and norm cap, and differ only in the per-waypoint weight. Δ∂\Delta_\partial: boundary excess in percentage points (generated minus target mass within 0.05 of the boundary). Safe and B-Safe in %.

Guidance weight1-D: W$_1\downarrow$SafeΔ∂\Delta_\partial2-D: SW$_1\downarrow$SafeΔ∂\Delta_\partialSeq.: SW$_1$ (10−310^{-3})SafeMaze Hard: SafeB-SafeRewardReward (safe)
Hinge, constant0.197 ± 0.00697.10.53 ± 0.050.1471 ± 0.000199.60.00 ± 0.025.1 ± 1.9100.013.512.087.48 ± 60.48116.50 ± 20.76
Hinge, λt=1/(2σt2)\lambda_t=1/(2\sigma_t^2)1.254 ± 0.01069.650.91 ± 0.250.506 ± 0.00791.415.47 ± 0.163.5 ± 2.3100.02.12.133.55 ± 55.44113.50 ± 8.96
ANCHOR, constant0.237 ± 0.002100.04.39 ± 0.170.184 ± 0.017100.00.62 ± 0.053.2 ± 1.7100.022.919.877.73 ± 60.05119.55 ± 22.11
ANCHOR, 1/σt1/\sigma_t0.059 ± 0.005100.0−0.33 ± 0.000.200 ± 0.018100.0−0.13 ± 0.0010.7 ± 1.899.945.335.473.33 ± 60.28118.77 ± 22.01

Projection baselines. We evaluated the released PDM implementation of Christopher et al. (2024) on the Gaussian mixtures, Sequential, and Multimodal Maze. It attains 100% safety on the mixtures with large distributional error (W$_1$ 2.887, SW$_1$ 1.968), 33.3% safety on Sequential and Maze Medium, and 0% on Maze Hard, at up to 55× the planning time of ANCHOR. We also tried the released DPCC implementation (Römer et al., 2025) with our models and constraints. Its dense trajectory-level optimization requires memory quadratic in the number of waypoints; the original experiments use 8 waypoints, whereas our maze benchmarks require 384–768. In our runs, the first five denoising-and-projection steps took 1.33–4.25 hours each and returned without convergence.

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

This paper proposes ANCHOR, a training-free guidance method for safety-constrained diffusion trajectory planning. Rather than ad-hoc locally projecting generated plans onto the feasible set, the proposed method uses an approximation of Doob's h-function to sample from the learned trajectory distribution, which is conditioned on safety. The paper introduces a probabilistic guidance approach to establish a theoretical safety guarantee under assumptions on obstacle geometry, the data distribution, and the diffusion process itself. Experiments on a wide variety of environments demonstrate improvements in distributional fidelity and planning performance, although practical safety and robustness remain mixed.

Critical strengths

(1) Structured, theory-based approach to safe trajectory planning. By formulating safe planning as sampling from p0(x0∣x0∈S)p_0(x_0 | x_0 \in S), the method aims to generate safe trajectories by almost construction, rather than simply ad-hoc correcting unsafe trajectories. The research question itself is not new, but in my opinion it is among the most critical and interesting ones in the current literature of generative AI-based safe planning. The proposed method itself is also interesting.

(2) Strong experimental evidence for distributional fidelity. The Gaussian-mixture experiments show that ANCHOR achieves substantially lower Wasserstein errors than the tested baselines while maintaining 100% observed safety. The results also demonstrate that ANCHOR better preserves the safety-conditioned distribution without excessively concentrating samples near obstacle boundaries, supporting the main motivation of the paper.

(3) Extensive experimental validation. ANCHOR was evaluated across a wide range of environment settings, including synthetic distributions, navigation, locomotion, and manipulation tasks. The results demonstrate promising improvements over existing methods, especially in multimodal navigation, and seem to support the practical effectiveness of the approach.

Critical weaknesses

(1) Limitations of the proposed guidance. Eq(18) approximates the joint safety probability using individual waypoint probabilities. Wouldn't that effectively neglect any correlations between the waypoints themselves? Then this approximation may not be generally valid for correlated trajectories, even when the posterior distribution is Gaussian. Although the synthetic experiments show promising results, further validation may be needed to establish whether these approximations remain reliable for more complicated, higher-dimensional trajectories.

(2) Lack of theoretical justification for distributional fidelity. While Theorem 2 establishes almost-sure safety under its assumptions, it does not guarantee that the resulting trajectories follow the desired conditional distribution p0(x0∣x0∈S)p_0(x_0 | x_0 \in S). In particular, the required guidance multiplier γ>1\gamma > 1 generally changes the sampling distribution relative to the exact Doob guidance. This looks to be an important gap between the theoretical result and the paper's central motive of preserving the learned distribution.

Questions for authors

(1) Waypoint correlations in eq(18). The discussion after Theorem 2 notes that correlations between waypoints can be captured by Ξ\Xi in eq(20). How does the guidance function in eq(18) account for these correlations when approximating the joint safety probability using individual waypoint probabilities? Could the authors clarify how this approximation affects the accuracy of the intended safety-conditioned distribution, even for strongly correlated trajectories?

(2) Safety versus distributional fidelity. Theorem 2 establishes almost-sure safety for a guidance multiplier γ>1\gamma > 1, whereas the exact Doob guidance in eqs.(6)–(8) is motivated by preserving the safety-conditioned distribution p0(x0∣x0∈S)p_0(x_0 | x_0 \in S). How does this additional guidance scaling affect distributional fidelity? Is this scaling is used in the distributional-fidelity experiments in Section 4.1? Any justification that the modified guidance can preserve the intended conditional distribution while guaranteeing safety?

(3) Relationship between theoretical assumptions and experiments. The paper's theory establishes safety under several simplifying assumptions, while the experiments in Sections 4.1–4.3 demonstrate promising results in more general settings. I understand Section 3.3 already acknowledges that extensions to more general obstacles and γ=1\gamma = 1 remain unproven. But could the authors more explicitly clarify which theoretical assumptions are relaxed in the experimental settings?

(4) Minor comments:

  • In Theorem 2, the almost-sure safety guarantee applies to guidance scaled by γ>1\gamma>1, but eq(19) is written without this multiplier.
  • Definition 1 assumes a closed obstacle set, but Theorem 2 defines an open half-space.
  • The citation to Ho & Salimans (2022) for classifier guidance in Section 4 appears incorrect, as this reference concerns classifier-free guidance. I think the authors may have meant Dhariwal & Nichol (2021)?
Reviewer_rrYHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a new planning method under hard safety constraints via safe diffusion by realzing the reverse process via a Doob h-transformation. The paper provides theoretical validity of the methods and runs an experimental evaluation with existing methods.

Critical strengths
  • The paper is mostly easy to read
  • The paper builds upon existing work
Critical weaknesses
  1. The paper has limited novelty over the work of Xiao et al 2023. The only evidence we have in the work is experimental tables which half of them show no significant improvement when accounting for the variance of the experiments.
  2. The experimental setup is extremely limited to toyish and overly synthetic scenarios, which makes it questionable whether this method scales to real planning tasks as claimed in the work.
  3. The method itself is rather restricted, eg. the half-space constraint, which often do not cover real settings.
Questions for authors

How does the method go beyond the limitations of Xiao et al. (2023) beyond parametrizing the CBF with a training-free analytical approximation?

Comments on the paper

2 comments
Area_Chair_V1CfOpenReview ↗

Dear Reviewers

Since the author-reviewer discussion ends soon, I encourage the reviewer to review the authors' rebuttals and let them know whether their responses are satisfactory and whether any concerns remain, if you haven't done so.

Thank you for your time and effort in reviewing for ICLR 2027.

Best wishes,

AC

We provide a summary of the rebuttal here. We have responded to all questions raised by the reviewers and substantially strengthened both the theoretical analysis and the experimental evaluation. These additions are incorporated into the updated manuscript (colored in blue) and summarized below.

Theory and Analysis

  • Stronger safety guarantees. We removed the requirement for additional guidance scaling, extended the theorem beyond half-spaces to smooth convex obstacles and more general correlation structures, and proved existence and uniqueness of the guided process rather than assuming it. See our responses to Reviewers mDHV and 9u41.

  • Quantitative analysis of waypoint correlations. We derived the general joint-safety formulation and added a new proposition bounding the error introduced by the waypoint-wise probability approximation, directly addressing the concern about neglected correlations. See our response to Reviewer mDHV and the additional analysis in Appendix D.4.

  • Robustness and numerical analysis. We expanded the analysis of covariance and gradient approximations, added experiments without the norm cap and with exact obstacle distances, and studied how the number of sampling steps affects safety and distributional accuracy. See our responses to Reviewers mDHV and 9u41.

Experiments and Baselines

  • Rejection-sampling baseline. We added rejection sampling across the benchmarks, evaluating distributional fidelity, safety, and computational cost to establish when repeated unguided sampling is competitive and when ANCHOR provides an advantage. See our response to Reviewer 9u41.

  • Training-free guidance baseline. We added DPS-style hinge guidance using the same model and obstacle information as ANCHOR. See our response to Reviewer 9u41.

  • Projection-based comparisons. We evaluated PDM on the synthetic and navigation benchmarks and documented the computational and convergence difficulties encountered when applying DPCC to long trajectories. See our response to Reviewer 9u41.

  • Original benchmark settings and execution safety. We added evaluations under the original Maze2D and locomotion constraints, together with an analysis of task performance and early termination that distinguishes successful safe execution from merely avoiding violations during an incomplete rollout. See our response to Reviewer 9u41.

  • Computational cost and experimental protocols. We reported wall-clock comparisons, including both gradient implementations of ANCHOR, and clarified validation-based parameter selection, baseline adaptations, weighting variants, and noise-schedule normalization. See our response to Reviewer 9u41.

We sincerely thank Reviewers 9u41 and mDHV for their detailed assessments. They highlighted concrete strengths of the paper, such as the highly relevant research question we pose (mDHV) and our obstacle avoidance via rerouting (9u41), and provided us with constructive criticism, which gave us the opportunity to improve our work.

During the rebuttal, Reviewer 9u41 explicitly acknowledged that the additional experiments clarify when ANCHOR helps and that the revised proof answers their questions about the guidance multiplier and the existence of the guided process. Reviewer mDHV acknowledged substantial theoretical improvements. Both retained their weak-acceptance recommendations; following our final response on waypoint correlations, mDHV did not specify whether or why that concern remained.

Reviewer rrYH misunderstood the method and did not respond to our rebuttal or either subsequent follow-up, leaving their methodological misunderstandings unaddressed.

Given the significant effort in time and computational resources (approximately 210 GPU hours of experiments in roughly one week), we remain concerned that the unchanged scores do not fully reflect the substantial additions made during the rebuttal.

We therefore respectfully ask the Area Chair, Senior Area Chair, and Program Chairs to take these additions and the discussion record into account when making the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.