acceptodds
Under review as a conference paper at ICLR 2027

PRESTO: Prioritized Reasoning for Safe Trajectory Optimization with Foundation Models

Abstract

Ensuring safety in open-world embodied systems is difficult not only because novel objects and hazards arise at test time, but because existing formal methods treat all safety constraints uniformly–leading to overly conservative or infeasible behavior in complex scenes. This paper introduces PRESTO, a framework that leverages multimodal large language models (MLLMs) to generate interpretable safety constraints directly from visual observations, and embeds that into a prioritized safe planner. Our framework enables (i) open-vocabulary safety reasoning over previously unseen objects, (ii) priority-aware safety constraints based on semantic risk, (iii) planner-agnostic deployment as a safety layer, and (iv) interpretable safety constraints. We evaluate PRESTO across autonomous driving and robotic manipulation tasks, reducing collision rates by 59.25% when applied as a safety layer to existing planners, while also generalizing to manipulation settings requiring semantic reasoning about fragile, hazardous, and low-risk objects. Our results highlight the potential of combining high-level multimodal reasoning with formal safety guarantees for scalable and interpretable autonomy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
safety, multimodal large language models, autonomous vehicles

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

This paper proposes PRESTO, a safety layer that combines multimodal language model reasoning with high-order control barrier functions. The model identifies objects, assesses their safety relevance, and assigns semantic priorities that guide constraint relaxation when modifying nominal control inputs. Validation includes driving and manipulation case studies, integration with existing planners on Dev10 and Bench2Drive, hardware demonstrations, and an ablation of semantic prioritization.

Critical strengths

The paper addresses a meaningful limitation of purely geometric safety constraints: treating obstacles uniformly can cause unnecessary conservatism. Using semantic priorities to guide constraint relaxation offers a useful connection between visual-language understanding and control. The experiments provide evidence of practical utility across driving and manipulation. Table 3 shows collision reductions when PRESTO is added to existing planners, while Table 4 isolates the benefit of semantic prioritization with the bounding-box source and CBF formulation fixed within each comparison.

Critical weaknesses
  1. The evidence establishing the method’s comparative benefits remains limited. Tables 1 and 2 contain only 12 driving and 10 manipulation scenarios, with insufficient detail on repeated trials and statistical uncertainty. The driving comparisons also omit recent vision-language driving methods, leaving PRESTO’s additional value for language-informed planners unclear. Although Table 3 provides broader evaluation, it does not fully resolve these questions.
  2. Robustness to dynamic obstacles and environmental variation is not sufficiently validated. Although Appendix A.1 describes constant-velocity prediction and Sections 4.2–4.3 mention moving pedestrians and vehicles, the dynamic-obstacle settings and their coverage in Tables 1 and 2 are not clearly documented. The paper also lacks a controlled evaluation of motion prediction and sensitivity to MLLM query frequency and obstacle speed. A controlled evaluation along these dimensions would help establish whether the safety improvements remain reliable as obstacle information becomes outdated. In addition, the qualitative examples and aggregate results do not establish robustness under different weather and lighting conditions, such as nighttime.

Overall, the approach is promising, but these gaps limit my confidence in the empirical conclusions. I would consider increasing my score if the authors address these concerns with convincing clarification and targeted evidence.

Questions for authors
  1. Comparative evidence: Could the authors provide a matched-protocol comparison with a representative recent vision-language driving method, such as SimLingo[1] or AutoVLA[2], either directly or with and without PRESTO?
  2. Environmental coverage: Could the authors clarify the composition of the test environments, particularly the weather and lighting conditions, and provide additional experiments under other conditions, such as nighttime or rain, to assess the method’s robustness?
  3. Dynamic sensitivity: Which case studies contain moving obstacles, and does simulation continue during MLLM inference? A targeted experiment varying query frequency and obstacle speed, while accounting for actual inference latency, would help establish when the safety improvements remain reliable.

[1] Renz et al. SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment. CVPR, 2025.

[2] Zhou et al. AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. NeurIPS, 2025.

3 comments
AuthorsOpenReview ↗

Thank you for your suggestions and constructive feedback!! Rebuttal by Authors [pt. 1]

We thank Reviewer CfYy for the suggestions on strengthening our empirical evaluation. In response, we added a matched-protocol comparison with SimLingo, repeated the case studies over four seeds with mean and standard deviation, and added randomized bird's-eye-view experiments, dynamic CARLA experiments, and lighting and weather experiments. We address each point below.

Comparative evidence: Could the authors provide a matched-protocol comparison with a representative recent vision-language driving method, such as SimLingo or AutoVLA, either directly or with and without PRESTO?

We thank the reviewer for this suggestion. We evaluated SimLingo under the same scenarios and four random seeds as our method. We also implemented AutoVLA, but its publicly released checkpoint is finetuned for NuPlan rather than CARLA, and it did not generalize successfully to our evaluation scenarios. We therefore exclude its result from Table 1.

We report the SimLingo results in Table 1. SimLingo achieves a collision rate of 54.17 ±\pm 8.33% and route completion of 67.86 ±\pm 5.03%, slightly outperforming TransFuser++ , but performing substantially worse than PRESTO with a collision rate of 6.25±4.176.25 \pm 4.17 and route completion of 97.92±2.4297.92\pm 2.42. We further examined SimLingo's language outputs and observed two failure modes. First, SimLingo was able to correctly identify the hazard but it still led to collisions. In the opening door scenarios, SimLingo recognizes the hazards and suggests changing lanes, but its planned maneuvers vary across frames, resulting in collisions in 3/8 opening-door runs. Second, SimLingo struggles with unfamiliar objects. It describes trash bags as "the black car ahead" and fails to identify the shopping cart, resulting in collisions in all evaluated runs of these scenarios. These results suggests that semantic reasoning alone does not guarantee safe control

The evidence establishing the method’s comparative benefits remains limited. Tables 1 and 2 contain only 12 driving and 10 manipulation scenarios, with insufficient detail on repeated trials and statistical uncertainty.

We thank the reviewer for raising this concern. We have repeated the comparisons in Tables 1 and 2 across four random seeds and updated the results to report mean and standard deviation. Beyond these case studies, our original manuscript also includes a Bench2Drive evaluation over 220 driving scenarios (Table 3).

To further strengthen the evaluation, we added additional experiments in bird's-eye-view space to isolate the contributions of the safety hierarchy and priority ranking. We generated 100 scenarios across five random seeds (500 scenarios in total), varying obstacle positions and safety priorities. In scenarios where a collision-free path exists (Table 7), PRESTO achieves zero critical collisions across all 500 trials (95% upper confidence bound: 0.6%) with 96.6% route completion. In comparison, treating all constraints as soft results in a 12.8% critical collision rate, while treating all constraints as hard reduces route completion to 69.6%. We have included the experiment results and setup details in Appendix A3.2.

We additionally introduced two dynamic driving experiments in CARLA with MLLM perception. In the controlled stopping experiment, we varied the initial distance to the lead vehicle and the ego-vehicle speed across 97 configurations, each evaluated over three seeds (Fig. 10). We also evaluated an asynchronous cut-in scenario across three seeds (Fig. 11). Both experiments examine collision avoidance with moving obstacles while accounting for measured inference latency. We have included the experiment results and setup details in Appendix A3.3.

AuthorsOpenReview ↗

Rebuttal by Authors (pt. 2)

Environmental coverage: Could the authors clarify the composition of the test environments, particularly the weather and lighting conditions, and provide additional experiments under other conditions, such as nighttime or rain, to assess the method’s robustness?

We thank the reviewer for raising this point. Our original case studies use the default weather settings in CARLA Town10HD. Separately, our Bench2Drive evaluation in Table 3 includes varying weather and lighting conditions. The Bench2Drive evaluation contains 27 weather configurations, including nighttime and rain, with their distribution shown in Figure 8 of [1].

To further assess robustness to environmental conditions, we extended four of our case-study scenarios to different lighting (sunset and nighttime) and weather (fog and rain) conditions, with three seeds per scenario. As shown in Table 14, PRESTO maintains a 0% collision rate and 100% route completion across all tested conditions. We also provide qualitative results in Figure 12. Under fog, the MLLM occasionally misidentifies objects, such as describing hay bales as “round objects,” while still correctly classifying them as collision-critical. Under rain, water droplets introduce local blur and partial occlusion in the camera images.

We have added these experimental results, implementation details, and additional analyses to Appendix A.3.4 ("Lighting and Weather").

Dynamic sensitivity: Which case studies contain moving obstacles, and does simulation continue during MLLM inference? A targeted experiment varying query frequency and obstacle speed, while accounting for actual inference latency, would help establish when the safety improvements remain reliable.

We thank the reviewer for raising these questions. Our original case studies consist of static obstacles, while the Bench2Drive evaluation includes moving traffic. Both use a synchronous simulation configuration, consistent with prior Bench2Drive evaluations [2]. In this setting, the simulation is paused and waits for MLLM inference and to complete. To assess the effect of inference latency in dynamic scenes, we added two experiments with moving vehicles under asynchronous execution.

First, we evaluated a controlled stopping scenario in which a lead vehicle brakes in front of the ego-vehicle. We varied the initial bumper-to-bumper distance and ego-vehicle speed from 10 to 30 km/h to examine how inference latency affects collision avoidance (Fig. 10). We compared MLLM perception, with a measured latency of 0.83 ± 0.07 s, against oracle perception with an latency of 0.04 s. The results quantify the additional stopping distance associated with MLLM inference latency, showing that higher speeds require larger initial gaps to avoid collisions. This suggests that the current implementation is more suitable for lower-speed operation.

Second, we evaluated PRESTO online and asynchronously in a dynamic cut-in scenario (Fig. 11). In this scenario, a sedan merges into the ego-vehicle’s lane and brakes abruptly. In this online setting, a new MLLM query is triggered whenever the previous inference completes, with a mean latency of 0.94 s. Across three seeds, the ego-vehicle completes the route without collision and maintains a minimum distance of 0.48 m from the sedan that cut-in front.

Together, these experiments characterize the operating conditions under which inference latency affects collision avoidance and demonstrate asynchronous execution in a dynamic driving scenario. We have included more analyses and experimental setup details in Appendix A.3.3.

References:

[1] Jia, X., Yang, Z., Li, Q., Zhang, Z. and Yan, J., 2024. Bench2Drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track.

[2] Renz, K., Chen, L., Arani, E. and Sinavski, O., 2025. SimLingo: Vision-only closed-loop autonomous driving with language-action alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.11993–12003.

Thank you to the authors for the detailed rebuttal and extensive additional experiments, which have substantially addressed my previous concerns. In particular, the additional comparisons, robustness evaluations under varying environmental conditions, and asynchronous experiments with dynamic obstacles provide stronger empirical support for the proposed approach.

These results further demonstrate the potential of combining semantic understanding with control-theoretic safety constraints, enabling the method to leverage semantic reasoning while providing meaningful safety benefits. I also agree with the authors that inference latency currently limits the method's applicability, particularly in higher-speed scenarios. Further reducing inference latency would improve its robustness and practical applicability in dynamic environments.

Overall, I believe the authors have adequately addressed my major concerns through their thorough responses and additional experiments. I have therefore decided to increase my rating.

Reviewer_yx2oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PRESTO uses a multimodal language model to infer object geometry, collision criticality, and semantic priority from visual observations. The inferred priorities determine slack penalties in a high-order control barrier function (HOCBF) quadratic program that modifies a nominal controller. The framework is evaluated in CARLA driving, simulated manipulation, and physical robot demonstrations. The main driving result reduces UniAD's collision rate from 49.08% to 20.00% on Bench2Drive, with a modest increase in route completion. Ablations with both predicted and oracle geometry indicate that semantic prioritization can improve progress in cluttered scenes.

Critical strengths
  1. Useful empirical evidence that semantic priorities can reduce excessive conservatism. Table 4 fixes the geometry source and CBF formulation within each comparison. With oracle boxes, semantic priority improves route completion from 82.16% to 100% while retaining zero reported collisions. This control helps separate the benefit of semantic prioritization from better perception and supports the paper's core practical motivation.

  2. Integration with existing planners and multiple embodied settings. Table 3 reports a substantial collision reduction on 220 Bench2Drive scenarios, from 49.08% to 20.00%, while route completion increases from 68.17% to 69.46%. Dev10 experiments use multiple planners, and manipulation examples include physical hardware. These results make the framework more compelling than a demonstration limited to a single toy scene.

Critical weaknesses
  1. The soft-constrained QP does not establish the stated hard safety or strict priority properties. Equation (3) permits slack in every safety-critical constraint with a finite quadratic penalty, whereas the preceding HOCBF condition requires a nonnegative barrier expression. Allowing negative slack can violate that condition even when a safe control exists. For example, for the first-order special case at a safety boundary, minimizing (u+1)2+pδ2(u+1)^2+p\delta^2 subject to u≥δu\geq\delta and u∈[−1,1]u\in[-1,1] gives u=δ=−1/(1+p)<0u=\delta=-1/(1+p)<0 for finite p>0p>0, although u=0u=0 is safe and feasible. Thus larger penalties alone do not guarantee safety. Similarly, a weighted sum of squared violations does not enforce the statement that higher-priority constraints are relaxed only after lower-priority ones (lines 236–238). Section 5 appropriately acknowledges perception uncertainty, but that qualification does not resolve this issue even with perfect perception. The paper needs an explicit hard-constraint subset or a justified hierarchical formulation, or must narrow its claims to empirically evaluated soft safety preferences.

  2. The closed-loop deployment evidence does not clearly connect the reported safety gains to the measured timing configuration. Section 3.2 specifies Qwen3-VL-4B, and Section 4.1 states an MLLM query interval of 0.75 seconds. However, Appendix A.3.2 reports about 10.16 seconds per Qwen3-VL-4B inference. The faster 0.915-second pipeline instead uses LFM2.5-VL with an additional 2D-to-3D lifting component and is evaluated on only four case-study scenarios. These may be legitimate different configurations, but the paper does not explain how the main Bench2Drive evaluation handles model latency, stale observations, and the safety-control update schedule. For moving hazards, this is consequential to whether the measured collision improvements support the proposed online safety layer. A configuration-to-result mapping and explicit simulation-time/wall-time protocol are needed.

Questions for authors
  1. Clarification/correction. What are the domain constraints on each slack variable in Equation (3), which safety-critical constraints are enforced without relaxation, and what theorem applies to the actual controller? How is the finite-penalty counterexample above excluded? Please also give the rank-to-weight mapping and penalty scale.
  2. Clarification. Is the intended priority semantics a weighted trade-off or a strict hierarchy? If the latter, what guarantees the claimed order of relaxation? How does the reported infeasibility at a 2.29 m gap (Table 8) arise if all obstacle constraints have unrestricted slack and the control-bound set is nonempty?
  3. Clarification. Which backbone, token budget, perception/lifting pipeline, and controller rate produced each main table? During CARLA evaluation, does simulation time pause during inference, and what information/control is used while waiting for a new prediction? Please explain the relationship between the 0.75-second query interval and Table 7's latency.
  4. Existing-result analysis. Can the authors report collision/progress outcomes and constraint violations from the main evaluation under the actual modeled observation delay, identifying which priority levels are relaxed? This would determine whether the claimed safety behavior holds in the evaluated operating regime.
10 comments
AuthorsOpenReview ↗

Thank you for your thorough and constructive feedback! Rebuttal by Authors (pt. 1)

We thank the Reviewer yx2o for the careful technical reading. The analysis of Eq. 3 identified a real error in our presentation, and the suggestions on hierarchical formulation and timing have improved the paper. We address each point below.

The soft-constrained QP does not establish the stated hard safety or strict priority properties. [...] Please also give the rank-to-weight mapping and penalty scale. [...] The paper needs an explicit hard-constraint subset or a justified hierarchical formulation, or must narrow its claims to empirically evaluated soft safety preferences.

We thank the reviewer for this careful analysis. We agree that Eq. (3), as originally written, does not guarantee satisfaction of a hard safety or a strict order of relaxation. We have adopted the reviewer's suggestion of an explicit hard-constraint subset and made three changes to Section 3.3. Our framework also allows for an alternative approach in which all constraints are initially treated as hard constraints. If the resulting QP is infeasible or prevents task completion, constraints can be progressively relaxed in order of increasing semantic priority, starting with the least important, until a feasible solution that permits task completion is found.

(1) Explicit hard-constraint subset. We revised Eq. (3) so that the slack enters HOCBF condition as:

Lfψi,m−1(x)+Lgψi,m−1(x)u+αm(ψi,m−1(x))+δi≥0,L_f\psi_{i,m-1}(x) + L_g\psi_{i,m-1}(x) u + \alpha_m(\psi_{i,m-1}(x)) + \delta_i \geq 0,

δi=0,  i∈S,δi≥0,  i∈S^,\delta_i = 0,\; i \in S, \qquad \delta_i \geq 0,\; i \in \hat{S},

with δi=0\delta_i=0 for i∈Si\in S and δi≥0\delta_i\geq 0 for i∈S^i\in\hat S. Here, SS denotes collision-critical objects, whose constraints cannot be relaxed, and S^=V∖S\hat S=\mathcal V\setminus S denotes non-critical objects. (2) The hierarchy is stated explicitly. The hierarchy enforces a hard distinction between collision-critical and non-critical constraints, while semantic ranking determines the relative penalties for relaxing non-critical constraints. We have removed the claim that finite penalties guarantee a strict order of relaxation, as the formulation instead implements a weighted trade-off among non-critical objects.

(3) The rank-to-weight mapping is given explicitly. We now specify pi=λwip_i=\lambda w_i, with λ=1000\lambda = 1000 in our experiments and wi=∣S^∣+1−ri∑j∈S^(∣S^∣+1−rj)w_i=\frac{|\hat S|+1-r_i}{\sum_{j\in\hat S}(|\hat S|+1-r_j)}, where rir_i is the MLLM predicted rank within S^\hat{S} (1 = highest priority).

We also clarified the scope of the safety guarantee. Hard HOCBF constraints cannot be relaxed by the QP, and their satisfaction provides the standard HOCBF forward-invariance guarantee under the required feasibility and modeling assumptions. When the QP becomes infeasible, a stopping fallback is applied, without claiming guaranteed collision avoidance in that case.

We further clarified the relationship between the formulation and the original experimental implementations. In the Bench2Drive evaluation (Table 3), all considered obstacles were treated as collision-critical, without priority-based relaxation. We have now rerun the case studies in Tables 1, 2, and 4 using the revised safety layer.

To further evaluate the safety hierarchy and priority ranking, we added randomized experiments in bird's-eye-view space to isolate the contributions of the safety hierarchy and priority ranking. We generated 100 scenarios across five random seeds (500 scenarios in total), varying obstacle positions and safety priorities. In scenarios where a collision-free path exists (Table 7), PRESTO achieves zero critical collisions across all 500 trials (95% upper confidence bound: 0.6%) with 96.696.6% route completion. Making every constraint soft with equal weights collides in 12.8±2.612.8 \pm 2.6% of these scenarios and making every constraint hard reduces the route completion to 69.6±6.669.6 \pm 6.6%. When no collision-free path to the goal exists (Table 8), PRESTO stops without collision in all 500 trials. We also evaluate priority ranking in Table 9, where lower-priority objects are violated more frequently than higher-priority objects (63.2% versus 2.6%). We have included the experiment results, analyses, and setup details in Appendix A3.2.

Table 7: Random scenes with a safe path. When a safe trajectory exists, treating all objects as hard constraints is overly conservative, while treating all as soft constraints compromises safety. Our hierarchy maintains safety without sacrificing route completion significantly.

MethodCollision Rate (%) ↓Collisions ↓Route Completion (%) ↑Avg. min. dist. (m)
No CBF61.0 ± 5.6305/500100.0 ± 0.0−0.42 ± 0.08
All hard0.8 ± 0.84/50069.6 ± 6.61.40 ± 0.33
All soft, equal weights12.8 ± 2.664/500100.0 ± 0.00.46 ± 0.07
Ours0.0 ± 0.00/50096.6 ± 2.10.62 ± 0.04

[continued in pt. 2]

AuthorsOpenReview ↗

Rebuttal by Authors (pt. 2)

Table 8: Random scenes with no safe path. When no safe trajectory exists, treating all objects as soft constraints leads to collisions with safety-critical objects, while our hierarchy safely stops the vehicle when no collision-free path exists.

MethodCollision Rate (%) ↓Collisions ↓Stopped w/o Collision (%) ↑Avg. min. dist. (≥ 0)
All hard0.0 ± 0.00/500100.0 ± 0.00.67 ± 0.12
All soft, equal weights82.6 ± 3.6413/5000.0 ± 0.0−0.54 ± 0.03
Ours0.0 ± 0.00/500100.0 ± 0.00.56 ± 0.02

What are the domain constraints on each slack variable in Equation (3), which safety-critical constraints are enforced without relaxation, and what theorem applies to the actual controller?

The slack domains are now explicitly stated in Eq. (3): δi=0\delta_i=0 for every collision-critical object i∈Si \in S, and δi≥0\delta_i \geq 0 for every non-critical object i∈S^i \in \hat{S}. All constraints in SS are therefore enforced without relaxation. The applicable result is the HOCBF forward-invariance theorem [4, Theorem 4]. Under the standard assumptions, including an initial state within the HOCBF admissible sets and continuous satisfaction of the HOCBF conditions, the safe set is forward invariant, such that hi(x(t))≥0h_i(x(t)) \geq 0 for all i∈Si\in S and t≥0t\geq 0. However, feasibility is not guaranteed under bounded control inputs, and this theoretical guarantee does not directly account for perception uncertainty or inference latency. We have clarified these conditions in the manuscript and described the stopping fallback applied when the QP becomes infeasible.

How is the finite-penalty counterexample above excluded?

We thank the Reviewer for this counterexample. The reviewer's counterexample arises because a finite slack penalty allows the optimizer to trade safety-critical constraint satisfaction against nominal control tracking. With δi=0\delta_i=0 for all i∈Si \in S, this trade-off is no longer possible for collision-critical objects. Slack is permitted only for non-critical constraints.

We further examined this issue empirically by varying the slack penalty λ\lambda from 1010 to 10510^5. In scenarios where no collision-free path to the goal exists, treating critical constraints as soft results in collision rates of 78–92% across the tested penalties. In contrast, the revised controller with hard critical constraints achieves zero collisions across the same tested penalty values.

Is the intended priority semantics a weighted trade-off or a strict hierarchy? If the latter, what guarantees the claimed order of relaxation? How does the reported infeasibility at a 2.29 m gap (Table 8) arise if all obstacle constraints have unrestricted slack and the control-bound set is nonempty?

Our priority structure has two levels. The distinction between collision-critical SS and non-critical (S^\hat{S}) objects is strict. Within S^\hat{S}, semantic priorities determine a weighted trade-off rather than a strict order of relaxation. Higher-priority objects receive larger slack penalties, encouraging the controller to relax lower-priority constraints first. Alternatively, hierarchical QP could strictly guarantee the claimed order of relaxation within S^\hat{S} [3]. We have corrected this claim in the manuscript and explicitly specified the rank-to-weight mapping with λ=1000\lambda=1000.

We also evaluated this behavior empirically. In scenarios where avoiding both non-critical objects is infeasible, PRESTO preferentially violates the lower-priority constraint: the higher-priority object is hit in 2.6±1.52.6 \pm 1.5% of trials, compared with 63.2±4.2%63.2\pm4.2\% for the lower-priority object (Table 9).

Regarding the narrow-gap experiment (Table 12 in the revised manuscript), both obstacles are classified as collision-critical, and their constraints therefore cannot be relaxed. At a gap width of 2.29 m, the gap is narrower than the vehicle width so no control input satisfies both hard HOCBF constraints, rendering the QP infeasible and triggering the stopping fallback. This behavior is consistent with the hard-constraint subset now specified in Eq. (3).

AuthorsOpenReview ↗

Rebuttal by Authors (pt.3)

Clarification on Runtime Configuration. Section 3.2 specifies Qwen3-VL-4B, and Section 4.1 states an MLLM query interval of 0.75 seconds. However, Appendix A.3.2 reports about 10.16 seconds per Qwen3-VL-4B inference. The faster 0.915-second pipeline instead uses LFM2.5-VL with an additional 2D-to-3D lifting component and is evaluated on only four case-study scenarios. A configuration-to-result mapping and explicit simulation-time/wall-time protocol are needed. Which backbone, token budget, perception/lifting pipeline, and controller rate produced each main table? During CARLA evaluation, does simulation time pause during inference, and what information/control is used while waiting for a new prediction? Please explain the relationship between the 0.75-second query interval and Table 7's latency.

We thank the reviewer for pointing out the need to clarify. For our main driving case studies and Bench2Drive evaluation (Tables 1, 3), we used Qwen3-VL-4B with a token budget of 384 tokens. We evaluated these experiments in CARLA using synchronous simulation with a fixed timestep of 0.05 s (20 Hz), consistent with the official Bench2Drive evaluation protocol [1, 2]. The MLLM is queried every 15 simulation ticks (0.75 s of simulation time), while the low-level PID controller updates at every tick using the latest trajectory. During MLLM inference and control computation, CARLA waits for the computation to finish before advancing to the next simulation tick. Thus, the 0.75 s query interval refers to simulation time rather than wall-clock time, while the measured 10.16 s is the actual inference latency of Qwen3-VL-4B. The main evaluations therefore do not account for the vehicle continuing to move during inference.

Separately, to investigate lower-latency alternatives, we first benchmarked different MLLM configurations using a single input image, with the inference times reported in Table 11. Based on these results, we selected LFM2.5-VL-1.6B with vLLM, a token budget of 256 tokens, and a 2D-to-3D lifting stage using Boxer. We then evaluated this optimized configuration on four CARLA case-study scenarios under the same synchronous simulation setup. Overall, the framework achieved an average latency of 0.915 s (Table 5). We have clarified the runtime results in Section 4.5 and clarified the different runtime configurations in Appendix A.1.

Existing-result analysis. Can the authors report collision/progress outcomes and constraint violations from the main evaluation under the actual modeled observation delay, identifying which priority levels are relaxed? This would determine whether the claimed safety behavior holds in the evaluated operating regime.

We thank the reviewer for raising this point. As clarified in our response above, the main CARLA and Bench2Drive evaluations use synchronous simulation. The collision and route-completion results in Tables 1 and 3 therefore do not account for vehicle motion during inference latency. To assess the effect of inference latency in dynamic scenes, we added two experiments with moving vehicles under asynchronous execution using the optimized LFM2.5-VL configuration.

First, we evaluated a controlled stopping scenario in which a lead vehicle brakes in front of the ego-vehicle. We varied the initial bumper-to-bumper distance and ego-vehicle speed from 10 to 30 km/h to examine how inference latency affects collision avoidance (Fig. 10). We compared MLLM perception, with a measured latency of 0.83 ± 0.07 s, against oracle perception with an latency of 0.04 s. The results quantify the additional stopping distance associated with MLLM inference latency, showing that higher speeds require larger initial gaps to avoid collisions. This suggests that the current implementation is more suitable for lower-speed operation.

Second, we evaluated PRESTO online and asynchronously in a dynamic cut-in scenario (Fig. 11). In this scenario, a sedan merges into the ego-vehicle’s lane and brakes abruptly. In this online setting, a new MLLM query is triggered whenever the previous inference completes, with a mean latency of 0.94 s \pm. Without the safety layer, the ego follows its nominal route and collides with the sedan after 4.4,s. With PRESTO, the ego-vehicle completes the route without collision and maintains a minimum distance of 0.48 m from the merging sedan (Fig. 11).

[continued in pt. 4]

Reviewer_qDooOpenReview ↗
Rating
3
Confidence
4(+2)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRESTO, a framework that uses MLLMs to generate interpretable, risk-prioritized safety constraints from visual observations and integrates them into existing planners. The method supports open-vocabulary safety reasoning and adapts to previously unseen objects. Experiments on autonomous driving and robotic manipulation tasks demonstrate improved safety.

Critical strengths
  • Clear and practical motivation. The paper addresses an important challenge in safety-critical embodied systems: balancing safety constraints with task feasibility. The motivation for introducing semantic safety priorities is intuitive and well justified.

  • Intuitive and interpretable framework. The proposed framework combines MLLM-based semantic reasoning with control barrier functions in a straightforward and interpretable manner. Its planner-agnostic design also makes it potentially applicable to different autonomous systems.

  • Comprehensive experimental evaluation. The authors evaluate the framework across autonomous driving and robotic manipulation tasks, including simulation benchmarks and real-world robotic experiments. The results demonstrate promising improvements in collision reduction and task completion.

Critical weaknesses
  • Unvalidated safety priority ranking.

    It is unclear whether a general-purpose MLLM without task-specific fine-tuning can reliably rank objects according to their safety importance. Although the authors demonstrate high accuracy in collision-critical classification, this does not necessarily indicate that the predicted safety priority rankings are reliable. The paper lacks a direct evaluation of ranking quality.

  • Inference latency concerns.

    I am concerned about the inference latency introduced by MLLMs in safety-critical applications such as autonomous driving and robotic manipulation. The reported MLLM inference time accounts for a substantial portion of the overall runtime, raising questions about the practicality of deploying this framework in real-time environments.

  • Unclear sources of performance improvement.

    It is difficult to determine whether the performance improvements primarily come from the MLLM-generated 3D bounding boxes, collision-critical classification, or safety priority ranking. More comprehensive ablation studies are needed to isolate the contributions of these components.

Questions for authors
  • Since the output of MLLM is in the structured json format. Is it possible to use a faster decision model to replace MLLM?

  • Figure 5 provides a case study. I'm not sure whether the right case is considered safe? It seems like there is still an overlap between the trajectory and the laptop.

6 comments
AuthorsOpenReview ↗

Thank you for your feedback! Rebuttal by Authors (pt. 1)

We thank Reviewer qDoo for the constructive feedback. The questions regarding ranking reliability led us to add a human ranking-agreement study, the questions regarding the sources of performance led to several ablations of the safety hierarchy and priority ranking, and the latency concerns motivated new asynchronous CARLA experiments. We address each point below.

Unvalidated safety priority ranking: It is unclear whether a general-purpose MLLM without task-specific fine-tuning can reliably rank objects according to their safety importance. Although the authors demonstrate high accuracy in collision-critical classification, this does not necessarily indicate that the predicted safety priority rankings are reliable. The paper lacks a direct evaluation of ranking quality.

We thank the reviewer for their careful consideration of this aspect of our method. As there is no established benchmark for evaluating semantic safety-priority rankings in these scenarios, we extended our human annotation study to directly assess agreement between MLLM-generated rankings and human judgments. Four human annotators independently ranked objects according to their safety importance in the same hardware scenarios. We compared the MLLM-generated rankings against the human annotations using Kendall's τ∈[−1,1]\tau \in [-1, 1]. Qwen2.5-VL-4B achieved a mean Kendall’s τ\tau of 0.644±0.2880.644 \pm 0.288, indicating substantial agreement with human rankings. We report this additional result in Table 6(c).

Inference latency concerns: I am concerned about the inference latency introduced by MLLMs in safety-critical applications such as autonomous driving and robotic manipulation. The reported MLLM inference time accounts for a substantial portion of the overall runtime, raising questions about the practicality of deploying this framework in real-time environments.

We thank the reviewer for raising this important concern. We acknowledge that MLLM inference latency remains a limitation for real-time deployment in safety-critical applications. In our main CARLA evaluations, the simulator runs synchronously—consistent with the official Bench2Drive evaluation protocol [1]—and waits for MLLM inference before advancing, so the reported inference latency represents wall-clock time rather than simulation time. In practice, our framework could operate alongside a high-frequency low-level safety controller that maintains safety while semantic reasoning is performed asynchronously. Although such a controller may conservatively stop the robot in challenging scenarios, our framework can provide semantic priorities that allow it to proceed once reasoning is complete. To further investigate the impact of inference latency, we conducted two additional experiments in CARLA with asynchronous execution. First, we evaluated a controlled stopping scenario in which a lead vehicle brakes in front of the ego-vehicle. We vary the initial bumper-to-bumper distance and ego-vehicle speed to examine how inference latency affects whether the ego-vehicle could avoid collision (Fig. 10). The results quantify the minimum stopping gaps required at different speeds, showing that higher speeds require larger initial gaps to avoid collisions. Second, we evaluate the framework online and asynchronously in a dynamic cut-in scenario where a sedan merges into the eg0-vehicle’s lane lane and brakes abruptly. The ego-vehicle completes the route without collision and maintains a minimum distance of 0.48 from the sedan (Fig. 11). Together, these experiments characterize the operating conditions under which inference latency affects collision avoidance and demonstrate asynchronous execution in a dynamic driving scenario. We have included more analyses and experimental setup details in Appendix A.3.3.

In addition to these new experiments, our original manuscript includes an inference latency study using smaller MLLMs (Table 5, Table 11). With an optimized LFM2.5-VL configuration using vLLM, we achieve an average end-to-end inference time of 0.915 s (1.13 Hz). Although this reduces latency, the current runtime still limits deployment in highly dynamic environments, a challenge also acknowledged in prior MLLM-based robotics work [2, 3]. Our additional experiments quantify the impact of this limitation on collision avoidance and demonstrate asynchronous operation in a low-speed driving scenario. We believe that continued advances in smaller, more capable VLMs [4, 5], including models designed for edge deployment such as LFM3-VL [6], together with further inference optimizations, will help reduce latency and improve the feasibility of real-time deployment.

AuthorsOpenReview ↗

Rebuttal by Authors (pt. 2)

Unclear sources of performance improvement: It is difficult to determine whether the performance improvements primarily come from the MLLM-generated 3D bounding boxes, collision-critical classification, or safety priority ranking. More comprehensive ablation studies are needed to isolate the contributions of these components.

We thank the reviewer for this suggestion. Our original manuscript included an ablation comparing MLLM-generated and oracle bounding boxes, as well as the effect of semantic prioritization (Table 4). We have now added further experiments to isolate the contributions of collision-critical classification and priority ranking. Figure 8 provides an overview of these ablations.

MLLM-generated 3D bounding boxes. In our original Table 4, we compared MLLM-generated and oracle bounding boxes while keeping semantic prioritization fixed. With oracle boxes, our method achieves 0% collisions and 100% route completion, compared with 6.25 ± 4.17% collisions and 97.92 ± 2.42% route completion using MLLM boxes. We further examine sensitivity to localization errors by perturbing oracle bounding boxes (Table 13). The method remains collision-free at 0.1 m and 5∘5^{\circ} perturbations, while collision rates increase to 8.33% at 0.3 m and 10∘10^{\circ}, and 33.33% at 0.5 m and 45∘45^{\circ}.

Collision-critical classification. We added experiments in bird's-eye-view space with known object geometry and criticality labels to isolate the contribution of the safety hierarchy. For each experimental setting, we generated 100 scenarios across five random seeds (500 trials), varying obstacle positions and safety priorities. When a safe path exists (Table 7), PRESTO achieves zero critical collisions across all 500 trials (95% upper confidence bound: 0.6%) with 96.6% route completion. In comparison, treating all constraints as soft results in a 12.8% critical collision rate, while treating all constraints as hard reduces route completion to 69.6%. When no safe path exists (Table 8), PRESTO avoids critical collisions in all 500 trials by stopping, whereas the all-soft baseline collides in 82.6% of trials. These results show that the safety hierarchy avoids critical collisions while allowing progress when strict avoidance of every object would be overly conservative.

Table 7: Random scenes with a safe path. When a safe trajectory exists, treating all objects as hard constraints is overly conservative, while treating all as soft constraints compromises safety. Our hierarchy maintains safety without sacrificing route completion significantly.

MethodCollision Rate (%) ↓Collisions ↓Route Completion (%) ↑Avg. min. dist. (m)
No CBF61.0 ± 5.6305/500100.0 ± 0.0−0.42 ± 0.08
All hard0.8 ± 0.84/50069.6 ± 6.61.40 ± 0.33
All soft, equal weights12.8 ± 2.664/500100.0 ± 0.00.46 ± 0.07
Ours0.0 ± 0.00/50096.6 ± 2.10.62 ± 0.04

Table 8: Random scenes with no safe path. When no safe trajectory exists, treating all objects as soft constraints leads to collisions with safety-critical objects, while our hierarchy safely stops the vehicle when no collision-free path exists.

MethodCollision Rate (%) ↓Collisions ↓Stopped w/o Collision (%) ↑Avg. min. dist. (≥ 0)
All hard0.0 ± 0.00/500100.0 ± 0.00.67 ± 0.12
All soft, equal weights82.6 ± 3.6413/5000.0 ± 0.0−0.54 ± 0.03
Ours0.0 ± 0.00/500100.0 ± 0.00.56 ± 0.02

[continued in pt. 3]

AuthorsOpenReview ↗

Rebuttal by Authors (pt. 3)

Priority ranking. To isolate the effect of rank ordering, we added a forced-choice experiment in which two non-critical objects cannot both be avoided (Table 9). Our method preferentially violates the lower-priority object, with collision rates of 63.2% for the lower-priority object compared with 2.6% for the higher-priority object. We further examine the effect of ranking under collision-criticality classification errors (Fig. 9). At a 50% misclassification probability, our approach reduces the critical collision rate from 36.8% to 4.2% by treating objects classified as non-critical as soft constraints rather than ignoring them entirely. When collision-critical objects are misclassified as non-critical with probability p=0.5p=0.5, removing constraints for objects labeled non-critical results in a 36.8% critical collision rate. In contrast, our approach maintains these constraints as priority-weighted soft constraints, reducing the collision rate to 4.2% (Fig. 9). Finally, to isolate the contribution of the MLLM-generated ranks, we replace them with random ranks, which increases the collision rate from 4.2% to 8.8% (Table 10). These results show that the MLLM-generated rankings improve collision avoidance and robustness to collision-criticality classification errors.

Table 9: Preferential violation of lower-priority constraints. When avoiding all objects is infeasible, our method preferentially violates lower-priority constraints.

MethodOverall Collision (%) ↓Overall Route Comp. (%) ↑High-Prior. Collision (%) ↓Low-Prior. Collision (%) ↑High-Prior. Min. dist. (m) ↑Low-Prior. Min. dist. (m) ↓
Ours0.0 ± 0.089.2 ± 2.42.6 ± 1.563.2 ± 4.20.82 ± 0.08−0.11 ± 0.04

Figure 9 (tabulated): Robustness to misclassification errors. Critical collision rate (%) when collision-critical objects are misclassified as non-critical with probability pp (500 trials per cell, mean ± std over 5 seeds).

Methodp=0p=0p=0.1p=0.1p=0.25p=0.25p=0.5p=0.5
Hierarchy, no ranking (drop non-critical)0.2 ± 0.48.4 ± 1.823.2 ± 5.936.8 ± 5.6
Ours (hierarchy + ranking)0.0 ± 0.01.4 ± 1.12.4 ± 1.94.2 ± 2.2

Table 10: Effect of rank ordering. Critical collision rate at p=0.5p=0.5.

MethodCollision (%) ↓
Random ranks8.8 ± 1.9
Ours4.2 ± 2.2

Since the output of MLLM is in the structured json format. Is it possible to use a faster decision model to replace MLLM?

We agree that a smaller, faster decision model could replace the MLLM. Our framework operates on structured object representations and does not depend on a specific MLLM backbone. We use an MLLM because it provides off-the-shelf semantic reasoning to identify collision-critical objects, assign safety priorities, and estimate 3D object bounding boxes without task-specific training, as demonstrated in Tables 1 and 2. A promising direction for future work is to distill these capabilities into a smaller model with priority prediction, which could reduce inference latency while retaining the same interface with our safety controller. However, whether such a model can retain the generalization capabilities of larger MLLMs, particularly for unseen objects and scenarios, would require further evaluation.

Figure 5 provides a case study. I'm not sure whether the right case is considered safe? It seems like there is still an overlap between the trajectory and the laptop.

We thank the reviewer for pointing out this ambiguity. The trajectory shown in Figure 5 represents the end-effector path. In this scenario, PRESTO considers the geometry of both the end-effector and the object it carries, with the objective of preventing the wet sponge from contacting the laptop. Therefore, the apparent overlap in the figure does not necessarily indicate a collision. We have clarified this distinction in the figure caption. While extending the safety constraints to the entire robot arm is possible, whole-arm collision avoidance is not considered in the current implementation and is left for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful and constructive feedback. We are encouraged to hear the reviewers,

  • Recognized the clear and practical motivation (Reviewers qDoo, yx2o) and the interpretable, convincing, and well-justified approach to integrating semantic reasoning with formal safety constraints (Reviewers CfYy, qDoo)
  • Highlighted the useful empirical evidence demonstrating how semantic prioritization can reduce excessive conservatism in geometric safety constraints (Reviewers yx2o, qDoo, CfYy).
  • Appreciated the comprehensive experimental evaluation across multiple embodied settings (Reviewers yx2o, qDoo, CfYy), including autonomous driving and robotic manipulation in simulation and on real hardware.
  • Highlighted the planner-agnostic nature of the framework (Reviewers qDoo, yx2o), enabling its integration with existing autonomy systems.

In response to the feedback, we have provided individual responses addressing each reviewer's remaining concerns. We have also updated the manuscript, with all changes highlighted in blue to improve clarity and address the reviewers' suggestions. Below, we summarize the additional experiments and revisions:

  • Conducted extensive ablation experiments to isolate the contributions of the safety hierarchy and priority ranking, demonstrating improved safety, reduced conservativeness, and robustness to classification errors. (Reviewers qDoo, yx2o, CfYy)
  • Added asynchronous CARLA experiments to evaluate inference latency in dynamic environments, quantitatively characterizing stopping distances across vehicle speeds and demonstrating collision avoidance in an online cut-in scenario. (Reviewers qDoo, yx2o, CfYy)
  • Improved the safety formulation based on Reviewer yx2o's feedback by explicitly distinguishing hard constraints for collision-critical objects from rank-weighted soft constraints for non-critical objects, which the reviewer subsequently acknowledged as resolved. This establishes conditional CBF safety guarantees for the collision-critical subset under the stated assumptions. We further evaluated this formulation empirically, achieving 0% critical collisions across 500 scenarios (95% upper confidence bound: 0.6%) while maintaining 96.6% route completion. (Reviewer yx2o)
  • Evaluated robustness under varying lighting and weather conditions, achieving collision-free performance across all tested scenarios. (Reviewer CfYy)
  • Extended the human evaluation study to directly assess MLLM-generated priority rankings using Kendall's τ, demonstrating substantial agreement with human judgments. (Reviewer qDoo)

For more details, please refer to our individual responses. We extend our gratitude to all reviewers for their time and valuable feedback. Please do not hesitate to share any further comments or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.