- 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 . The exact reverse process for this target is the Doob -transform, with guidance and . ANCHOR approximates in closed form in two steps. (i) It approximates by a Gaussian with the Tweedie mean and, in practice, the isotropic covariance . (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 . The guidance is over waypoints, and it is active only for . 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 with .
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
-
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.
-
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 -transform stays safe under a bounded tilt of the data distribution.
Critical weaknesses
-
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, , and it assumes that the guided SDE has a solution.
- All experiments use , and the authors note that the proof gives nothing for . They also use curved obstacles, a learned score with 256 discrete steps, and, on Maze, a norm cap of 2 on . 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 the guidance also departs from the Doob transform, so the theorem is about safety only, not about recovering . 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.
-
A key baseline is missing, and the tuning protocol is unclear.
- Rejection sampling from the unguided planner samples 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 (, 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 ) but do not seem to get similar tuning.
Questions for authors
Clarifications (could change my assessment):
- Were ANCHOR's hyperparameters (, , 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?
- What is the wall-clock planning time compared with Diffuser, SafeDiffuser, and SafeFlowMatcher, for both the through-model and stopped-model gradients?
- Does any safety argument survive the norm cap used on Maze? Can you show, even only empirically, how safety changes with and with the number of sampling steps?
- Theorem 2 assumes that the guided SDE has a solution with continuous paths. Is this guaranteed for the proposed drift, which grows like inside the obstacle?
- Which noise schedule and time normalization do the experiments use? does not match the OU process of Eq. (1) with large .
Experiments that would materially strengthen the paper:
- A rejection-sampling baseline (unguided Diffuser plus a safety check, draws) at matched compute on all benchmarks, including W1/SW1 on the synthetic tasks.
- Results in the original constraint settings (reference Maze2D, Hopper 1.5 m, Walker2d 1.3 m), where the safe conditional has no data support.
- 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 (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 .
| Method | Hopper 1.6 m: Safe (%) | Score (safe) | Score (unsafe) | Early (%) | Walker2d 1.4 m: Safe (%) | Score (safe) | Score (unsafe) | Early (%) |
|---|---|---|---|---|---|---|---|---|
| Diffuser | 36 | 0.382 | 0.455 | 72.2 | 67 | 0.137 | 0.353 | 89.6 |
| CG | 35 | 0.374 | 0.464 | 74.3 | 71 | 0.158 | 0.350 | 85.9 |
| Truncate | 39 | 0.367 | 0.457 | 59.0 | 73 | 0.152 | 0.365 | 87.7 |
| SafeDiffuser–RoS | 42 | 0.372 | 0.454 | 64.3 | 70 | 0.164 | 0.342 | 88.6 |
| Flow Matching | 67 | 0.323 | 0.448 | 91.0 | 100 | 0.105 | – | 98.0 |
| FlowMatcher | 97 | 0.134 | 0.351 | 100 | 97 | 0.051 | 0.258 | 97.9 |
| SafeFlowMatcher | 96 | 0.134 | 0.349 | 100 | 95 | 0.050 | 0.199 | 100 |
| ANCHOR (constant) | 41 | 0.362 | 0.463 | 73.2 | 72 | 0.163 | 0.357 | 84.7 |
| ANCHOR (inverse std) | 39 | 0.406 | 0.457 | 56.4 | 60 | 0.135 | 0.362 | 90.0 |
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 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.
| Method | Maze2D: Reward | Safe (%) | Hopper 1.5 m: Score | Safe P/E | Walker2d 1.3 m: Score | Safe P/E |
|---|---|---|---|---|---|---|
| Diffuser | 413.56 ± 77.87 | 1.6 | 0.431 ± 0.098 | 1/10 | 0.220 ± 0.151 | 0/49 |
| SafeDiffuser–RoS | 381.86 ± 117.22 | 99.0 | 0.422 ± 0.101 | 100/15 | 0.217 ± 0.149 | 100/53 |
| SafeFlowMatcher | 438.75 ± 2.59 | 100.0 | 0.130 ± 0.061 | 100/58 | 0.099 ± 0.051 | 100/8 |
| ANCHOR () | 407.22 ± 97.01 | 100.0 | 0.413 ± 0.105 | 100/14 | 0.228 ± 0.156 | 100/47 |
| ANCHOR () | 159.45 ± 96.46 | 100.0 | 0.125 ± 0.020 | 100/100 | −0.008 ± 0.004 | 100/97 |
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 unguided draws. As it samples 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 in the unguided acceptance rate , and its latency is random. It beats ANCHOR on Maze Medium (; 99.0% vs. 94.3%). On Maze Hard (), it falls behind (29.7% vs. 45.3%) at 5× the time. Under the original constraints (), 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 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
(Chung et al., 2023; Bansal et al., 2023), using the same Tweedie mean and SDF as ANCHOR. Inside the obstacle (), ANCHOR's guidance approximately equals the hinge gradient with step . 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 , so its time integral diverges as . 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, 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, and norm cap, and differ only in the per-waypoint weight. : boundary excess in percentage points (generated minus target mass within 0.05 of the boundary). Safe and B-Safe in %.
| Guidance weight | 1-D: W$_1\downarrow$ | Safe | 2-D: SW$_1\downarrow$ | Safe | Seq.: SW$_1$ () | Safe | Maze Hard: Safe | B-Safe | Reward | Reward (safe) | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Hinge, constant | 0.197 ± 0.006 | 97.1 | 0.53 ± 0.05 | 0.1471 ± 0.0001 | 99.6 | 0.00 ± 0.02 | 5.1 ± 1.9 | 100.0 | 13.5 | 12.0 | 87.48 ± 60.48 | 116.50 ± 20.76 |
| Hinge, | 1.254 ± 0.010 | 69.6 | 50.91 ± 0.25 | 0.506 ± 0.007 | 91.4 | 15.47 ± 0.16 | 3.5 ± 2.3 | 100.0 | 2.1 | 2.1 | 33.55 ± 55.44 | 113.50 ± 8.96 |
| ANCHOR, constant | 0.237 ± 0.002 | 100.0 | 4.39 ± 0.17 | 0.184 ± 0.017 | 100.0 | 0.62 ± 0.05 | 3.2 ± 1.7 | 100.0 | 22.9 | 19.8 | 77.73 ± 60.05 | 119.55 ± 22.11 |
| ANCHOR, | 0.059 ± 0.005 | 100.0 | −0.33 ± 0.00 | 0.200 ± 0.018 | 100.0 | −0.13 ± 0.00 | 10.7 ± 1.8 | 99.9 | 45.3 | 35.4 | 73.33 ± 60.28 | 118.77 ± 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.
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:
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:
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).