{"id":"6ece0b51-eeee-4adb-935f-60252cbb8e2d","arxiv_id":"2502.01924","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DualGuard-MPPI filters every sampled control sequence with a Hamilton-Jacobi safety filter, producing all-safe rollouts and empirically better performance than existing MPPI methods.","lead":"This paper proposes DualGuard-MPPI, a controller that adds Hamilton-Jacobi safety filtering inside the sampling loop of MPPI, a sampling-based model predictive control method, so every candidate trajectory is safe. It reports lower failure rates and better performance than prior MPPI variants in simulations and on a real RC car.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'provably safe' claim rests on a discrete-time least-restrictive filter with no margin analysis; inter-sample dynamics and numerical value-function error are not shown to be absorbed.","rationale":"I read the paper as a constructive integration of HJ reachability with MPPI. The experiments are well-designed, and the performance improvements over the LRF baselines are plausible because safe rollouts reduce sample wastage. The load-bearing vulnerability is the gap between the continuous-time safety proof of the least restrictive filter and the discrete-time numerical implementation used inside MPPI and at execution. This is the same weakness the reader identified: the 'provably safe' wording is stronger than the implementation supports unless one adds an explicit margin or a discrete-time invariance analysis. My proposed dense re-simulation would empirically stress that gap. If it reveals boundary crossings or value-function dips between control steps, the central safety claim collapses and conditional acceptance would be too weak. If it does not, the practical evidence is strengthened, but the paper should still state the margin explicitly and ideally provide a formal discrete-time argument. Since the reader's verdict is already CONDITIONAL and my concern does not change that level of acceptance, the appropriate label is UNCHANGED.","tokens_in":12109,"tokens_out":4901,"duration_ms":55562,"concrete_test":"Dense re-simulation test: for a subset of episodes from Secs. V-A and VI, run the closed-loop DualGuard system but integrate the dynamics (including the same LRF decisions made at the original controller rate) with m=10 substeps per control period. Record min V(x(τ)) and l(x(τ)) over all substeps. If any execution has min V < 0 or l ≤ 0 during an interval while the original-rate run reports failure=0, the discrete-time filter has no safety margin over one control period. A complementary analytical test: compute the maximum decrease of the numerical value function under the unmodified nominal controller over one Δt for the set {x: 0 < V(x) < ε} and compare it with the solver's local error estimate; if the decrease exceeds the error estimate, a positive switching margin ε is needed to support the guarantee.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim ('ensures that all generated samples are provably safe for the system') is inherited from the continuous-time least restrictive filter in eq. (11): apply unom while V(x)>0 and switch to u_safe(x) when V(x)=0. In the algorithm, however, this filter is evaluated only at discrete instants jΔt along each rollout (Alg. 1) and at execution time. For nonlinear dynamics integrated over a nonzero Δt, V(x_j)>0 at the start of an interval does not guarantee that the interval trajectory remains outside the BRT: the state can cross V=0 or enter the failure set before the next filter evaluation. The value function itself is a numerical solution of the HJI-VI on a grid (LevelSetToolbox), so V(x)>0 at a grid point or state is not proof that the true value is positive. No safety margin ε (e.g., switching when V(x) ≤ ε) or discrete-time invariance argument is provided, nor is the bounded disturbance in eq. (15) analyzed over a control period. Section VII lists offline BRT and explicit-dynamics limitations but not this gap. Therefore the hard-safety guarantee as stated is stronger than the implementation supports.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DualGuard-MPPI, a sampling-based model predictive control method that integrates Hamilton-Jacobi reachability analysis into the MPPI sampling procedure. An offline-computed backward reachable tube and value function are used to build a least-restrictive safety filter, which is applied both to each sampled rollout during the MPPI update and to the final control sequence at execution. The authors claim that this architecture ensures all generated samples are provably safe, eliminates the need for safety penalty tuning, improves sample efficiency, and outperforms penalty-based and single-filter MPPI baselines. The method is evaluated in two simulation domains (planar navigation and quadrotor navigation) and on a hardware RC car, with results reported over 100 episodes per simulation setting and a multiple-lap hardware run.","tokens_in":12393,"tokens_out":6382,"duration_ms":61049,"significance":"If the safety guarantee is made rigorous or appropriately qualified, the paper makes a useful contribution: it demonstrates a practical way to combine a least-restrictive filter with MPPI so that all samples contribute to the performance objective rather than being discarded as unsafe. The multimodality motivation for an output filtering stage is well presented, and the empirical evaluation is comparatively thorough: 100-episode simulations with standard errors and paired t-tests, three experimental settings, and a hardware demonstration. The performance improvements over penalty-based and single-layer-filter baselines are consistent and generally statistically significant. The main weakness is that the core safety guarantee is inherited from prior work [6] and is not re-derived or verified for the discrete-time, numerically approximated implementation used in the algorithm, which is a load-bearing issue for the paper's central claim.","major_comments":[{"comment":"The central safety claim is not established for the implemented discrete-time filter. The LRF guarantee in eq. (11) is a continuous-time statement: applying u_safe when V(x)=0 keeps the trajectory in the safe set. In Alg. 1, however, the filter is evaluated only at discrete instants along each rollout and at execution. For a nonzero interval Δt, V(x_j)>0 at the start of an interval does not prevent the state from crossing V=0 and entering the failure set before the next evaluation. The paper does not provide a margin ε (e.g., switching when V(x)≤ε) or a discrete-time invariance argument, nor does it analyze how the numerical approximation error in V(x) from the LevelSetToolbox is absorbed. Consequently, the statements 'ensures that all generated samples are provably safe' (abstract) and 'guaranteed to maintain safety' (§IV-A) are stronger than what the implementation supports. This gap is load-bearing for the main contribution, so it should be addressed by adding a margin analysis, a discrete-time certificate, or by explicitly qualifying the safety guarantee to the continuous-time idealization.","section":"§IV-A, eq. (11)"},{"comment":"The quadrotor experiment uses zero-mean Gaussian noise to simulate actuation errors, but the reachability-based guarantee requires bounded disturbances, as stated in §II and used in the BRT computation (eqs. (6)–(8)). If the Gaussian noise is unbounded, or if the BRT was computed with a different disturbance model than the one used in the simulation, the 'provably safe' claim does not apply to this experiment. The authors should either include the noise as a bounded disturbance in the reachability computation and report the bound, or state that the guarantee holds only for the nominal noiseless dynamics and that the Gaussian noise is an additional empirical robustness check.","section":"§V-B, eq. (13)"},{"comment":"The description of the output filtering stage is ambiguous with respect to the MPPI update rule. The text states that the filtered control perturbations Δk_j are used in the update rule (3), but (3) is defined in terms of the original random perturbations δk_j. If the filtered perturbations are substituted, the information-theoretic justification of the MPPI update no longer applies, and the effects on the optimality of the resulting controller should be discussed. In addition, the output LRF is applied only to the first control u*_0; since this filter is evaluated at a discrete instant, the same inter-sample safety issue raised for the rollouts applies to the executed trajectory as well.","section":"§IV-B"}],"minor_comments":[{"comment":"The sentence 'all methods with an output LRF avoid collisions as expected' reports an empirical observation over 100 episodes; this should be phrased as an empirical result, not as evidence for the theoretical guarantee.","section":"§V-A, Table I"},{"comment":"The pseudocode for Alg. 1 is referenced but does not appear in the manuscript; please include it or a more detailed algorithmic description, since the distinction between filtering at rollout time and at output time is central to the paper.","section":"Alg. 1"},{"comment":"RelCost is normalized to the proposed method's mean cost and computed only over trajectories that are common and safe for both methods; the number of common episodes is not reported, which makes the magnitude of the reported differences harder to interpret.","section":"§V, RelCost metric"},{"comment":"The running-cost expressions mix a continuous-time formulation in eq. (1) with discrete-time sums in eqs. (12) and (16); the discretization used in the implementation should be stated explicitly.","section":"§V, eqs. (12) and (16)"},{"comment":"The phrase 'provably safe' is used in the abstract and contributions; consider replacing it with a more qualified statement such as 'safe when the continuous-time LRF with the numerically computed value function is available' to avoid overclaiming in light of the discrete-time implementation.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The safety guarantee is cited to the authors' own prior work [6], which is peer-reviewed and appropriate; nevertheless, the current manuscript should more explicitly restate the assumptions under which that guarantee holds, since the discrete-time implementation is not covered. The empirical evaluation is solid and the paper is within the scope of eess.SY. The main revision needed is to close or clearly qualify the gap between the continuous-time LRF guarantee and the discrete-time algorithm."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know that the core idea here is sensible and the experimental work is above average for this subfield, but the paper overstates its safety guarantee. The authors combine two known tools—HJ reachability and MPPI—by applying a least-restrictive filter to each sampled rollout and then again to the output control. That two-stage composition is genuinely new relative to Shield MPPI and the filter-only baselines, and it fixes a real problem: unsafe samples waste computation and don't contribute to the update. Their results across the planar, quadrotor, and RC-car experiments are consistent and better than the baselines, with zero failures for all filtered methods and statistically significant cost improvements. The paired t-tests and the varying sample-count sweep are good practice. The soft spot is the \"provably safe\" claim. The LRF in eq. (11) is a continuous-time construction: apply u_nom while V(x)>0, switch at V(x)=0. The algorithm evaluates this filter only at discrete instants along each rollout and at execution time. For nonlinear dynamics integrated over a nonzero Δt, the state can cross the BRT boundary between evaluations. The value function is also numerical, so V(x)>0 at a grid point isn't a certificate. The paper doesn't provide a margin (e.g., switching at V(x) ≤ ε) or a discrete-time invariance argument, and it doesn't analyze how the bounded disturbances in eq. (15) are absorbed over a control period. So the statement that \"all generated samples are provably safe\" is stronger than the implementation supports. The stress-test note is on target here. This isn't fatal to the approach—in practice the filters likely work because the time steps are small—but it's a gap between the theorem and the algorithm. Two smaller things: no code or data release, so the empirical results are difficult to reproduce exactly. And the performance improvement is empirical, not explained analytically; that's fine, but it means the headline claim about sample efficiency rests on experiments rather than a mechanism proof. Who is this for? Researchers working on safe sampling-based MPC, especially with HJ reachability. It deserves a serious referee. The authors need to either soften the claim or add a margin analysis / discrete-time guarantee. I'd recommend major revision with a specific request for that analysis.","headline":"Clean algorithm and strong experiments, but the discrete-time safety guarantee is oversold.","tokens_in":12876,"tokens_out":2525,"would_cite":true,"duration_ms":25091,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49L25","93B03","93C85"],"pacs":[],"model":"deepseek-v4-flash","headline":"DualGuard-MPPI integrates Hamilton-Jacobi reachability filters into both the sampling and execution stages of MPPI, making every sampled rollout provably safe while improving performance.","keywords":["model predictive path integral control","Hamilton-Jacobi reachability","backward reachable tube","least restrictive filter","safe optimal control","sampling-based MPC","collision avoidance","safe exploration"],"falsifier":"Run DualGuard-MPPI with a value function computed on a grid matching the control period, and inject the worst-case disturbance from the dynamics model (for the RC car, $d_x,d_y=\\pm0.1$ m/s) starting from a state just inside the safe set; if any filtered rollout or executed trajectory reaches a state with value function $\\leq 0$ or enters the failure set, the paper's claim that all samples are provably safe is falsified, while a sweep of such adversarial initial states that always keeps $V(x)>0$ would support the claim.","tokens_in":11905,"feed_emoji":"🛡️","tokens_out":11161,"duration_ms":96792,"temperature":0.7,"pith_summary":"Sampling-based model predictive control, such as MPPI, handles complex dynamics but usually treats safety as a soft penalty, wasting samples and offering no guarantee. This paper proposes DualGuard-MPPI, which places a Hamilton-Jacobi reachability safety filter inside the MPPI sampling loop: every sampled control perturbation is filtered before its rollout is simulated, so all rollouts are provably collision-free and contribute to the performance update. A second filter is applied to the final control sequence to prevent unsafe averaging of distinct safe evasive maneuvers. The result is a controller that provably avoids failure while improving sample efficiency, success rate, and cost in planar navigation, quadrotor, and RC-car hardware experiments.","feed_headline":"Safety filters inside MPPI keep every sample collision-free","feed_subtitle":"Filtering every sampled rollout and the final command avoids collisions without safety penalty tuning.","key_machinery":"The key machinery is the least restrictive filter (LRF) built from a time-converged Hamilton-Jacobi reachability value function. The value function satisfies a Hamilton-Jacobi-Isaacs variational inequality, and its sub-zero level set is the backward reachable tube (BRT): the set of states from which a failure is unavoidable despite best control. The LRF switches between a nominal control and the reachability safe control exactly when the value function is zero, which keeps the value non-negative and the state inside the safe set. DualGuard-MPPI applies this LRF at every step of every sampled rollout during MPPI's forward simulation, and once more to the output control sequence, which is what makes every generated sample provably safe and prevents a multimodal mixture of safe but divergent maneuvers from averaging into an unsafe command.","core_discovery":"The paper's central claim is that safety constraints can be moved from the cost function into the sampling process itself, and that doing so is both safer and more performant. DualGuard-MPPI computes the Hamilton-Jacobi backward reachable tube of the failure set offline, giving a value function whose sub-zero level set marks states from which entering the failure set is unavoidable under worst-case disturbance. During each MPPI iteration, every sampled control sequence is passed through a least restrictive filter at every horizon step, so the simulated rollout is safe by construction; the filtered perturbation, not the raw sample, is used in the information-theoretic update. The resulting optimal control sequence is filtered once more before execution. Because all K samples are safe, none are discarded for safety reasons, effective sampling variance drops, and the remaining computation is spent optimizing performance. The paper reports zero safety violations and lower normalized cost across 100 random episodes in two simulation domains, and higher lap speed on a hardware RC car, relative to penalty-based, BRT-penalty, LRF-only, and Shield-MPPI baselines.","pith_inferences":["The same two-filter architecture could be applied to other sampling-based optimal controllers that aggregate weighted trajectories, if they have access to a similar value-function representation of the safe set.","If the offline BRT computation is replaced by online or learning-based reachability updates, the framework could extend to changing environments, though the safety guarantee would then inherit the approximation error of the updated value function.","The hardware results suggest that the in-sampling filter (not the output filter) provides most of the performance gain; a controlled ablation separating the two stages would quantify each stage's contribution.","A formal account of how the value-function grid resolution, control period, and disturbance bound interact would turn the empirical safety margin into a tunable design parameter, allowing users to trade conservatism for agility."],"forward_implications":["DualGuard-MPPI achieves zero safety violations in all tested domains while maintaining or improving success rate and cost relative to penalty-based MPPI and single-layer LRF baselines.","Because every sample is safe, the method sustains high performance with far fewer samples: with 60 samples it keeps a 96% success rate in the planar navigation domain, versus at most 70% for the closest baseline.","Safety penalty terms can be removed from the MPPI cost function, eliminating the need to tune safety-related weights.","The second filtering stage resolves the multimodality failure mode in which averaging two safe swerves (left and right around an obstacle) would produce a straight, unsafe command.","The full algorithm runs on an RC car at 50 Hz with 1000 parallel rollouts, staying within the 20 ms control budget."],"supporting_citations":[{"why":"Supplies the original MPPI sampling-based framework whose weighted update law this paper extends.","marker":"[5]"},{"why":"Provides the parallel-computable MPPI theory and the exact weighted-average update rule used by DualGuard-MPPI.","marker":"[8]"},{"why":"Defines the least restrictive filter and proves the safety guarantee that DualGuard's two filtering stages rely on.","marker":"[6]"},{"why":"Formulates the Hamilton-Jacobi-Isaacs value function and non-anticipative disturbance strategies used to characterize the safe set.","marker":"[22]"},{"why":"Presents the time-dependent HJI-VI formulation whose converged solution yields the backward reachable tube.","marker":"[24]"},{"why":"Supplies the numerical level-set solver used to compute the value function in all simulations and hardware experiments.","marker":"[25]"},{"why":"Defines Shield-MPPI, the state-of-the-art safety-constrained MPPI baseline that DualGuard is compared against.","marker":"[9]"}],"fun_headline_variants":["MPPI samples filtered by HJ reachability: safe and faster","DualGuard-MPPI: every sample provably safe, more performance","HJ safety filter inside MPPI beats penalty methods","Sampling-based MPC that never violates safety","Reachability-aware MPPI: zero collisions, optimal performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The safety guarantee rests on the assumption that the discrete-time safety check, made with a numerically computed map of unsafe states, acts often enough that the real system, with its estimation errors and disturbances, can never reach an unsafe state between two checks.","fun_headline_variants_meta":{"raw":{"variants":["MPPI samples filtered by HJ reachability: safe and faster","DualGuard-MPPI: every sample provably safe, more performance","HJ safety filter inside MPPI beats penalty methods","Sampling-based MPC that never violates safety","Reachability-aware MPPI: zero collisions, optimal performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2023,"prompt_tokens":953,"completion_tokens":1070,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":986}},"tokens_in":569,"tokens_out":1070,"duration_ms":10330,"temperature":1.0,"reasoning_tokens":986,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:57:45.688608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DualGuard-MPPI with a value function computed on a grid matching the control period, and inject the worst-case disturbance from the dynamics model (for the RC car, $d_x,d_y=\\pm0.1$ m/s) starting from a state just inside the safe set; if any filtered rollout or executed trajectory reaches a state with value function $\\leq 0$ or enters the failure set, the paper's claim that all samples are provably safe is falsified, while a sweep of such adversarial initial states that always keeps $V(x)>0$ would support the claim.","supporting_citations":[{"cited_title":"Information-theoretic model predictive control: Theory and applications to autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the original MPPI sampling-based framework whose weighted update law this paper extends."},{"cited_title":"Model predictive path integral control: From theory to parallel computation,","cited_arxiv_id":null,"evidence_quote":"Provides the parallel-computable MPPI theory and the exact weighted-average update rule used by DualGuard-MPPI."},{"cited_title":"Hamilton-Jacobi Reachability: A brief overview and recent advances,","cited_arxiv_id":null,"evidence_quote":"Formulates the Hamilton-Jacobi-Isaacs value function and non-anticipative disturbance strategies used to characterize the safe set."},{"cited_title":"A time-dependent Hamilton- Jacobi formulation of reachable sets for continuous dynamic games,","cited_arxiv_id":null,"evidence_quote":"Presents the time-dependent HJI-VI formulation whose converged solution yields the backward reachable tube."},{"cited_title":"A toolbox of level set methods,","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical level-set solver used to compute the value function in all simulations and hardware experiments."},{"cited_title":"Shield mppi: A compu- tationally efficient robust mpc method using control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Defines Shield-MPPI, the state-of-the-art safety-constrained MPPI baseline that DualGuard is compared against."}],"review_version":1}