{"id":"51627b98-6e29-4508-a442-0b6d742533cb","arxiv_id":"2604.12149","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"UGE-TO models trajectories as uncertainty-induced distributions and uses Hellinger distance to enforce sample diversity, yielding faster convergence and better success rates in sampling-based MPC.","lead":"The paper presents UGE-TO, which represents candidate trajectories as probability distributions shaped by uncertainty ellipsoids and separates them using Hellinger distance to improve coverage of the configuration space. This is integrated into UGE-MPC for sampling-based model predictive control, showing faster convergence and higher success rates than baselines in simulation and real-world tests.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Ellipsoid-induced distributions may not accurately capture nonlinear dynamics, so Hellinger separation need not yield better config-space coverage than action sampling.","rationale":"The reader's weakest assumption identifies precisely the same link; the concrete test above would falsify or corroborate it directly. No stronger internal inconsistency appears once the approximation quality is checked.","tokens_in":1805,"tokens_out":282,"duration_ms":37762,"concrete_test":"Replace the ellipsoid propagation inside UGE-TO with a 500-particle Monte-Carlo rollout using the same nominal controls, recompute the Hellinger-based selection on the empirical distributions, and re-run the cluttered-environment trials; if success-rate improvement over the best baseline falls below 3 percentage points, the distributional claim is sensitive to the approximation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The method represents trajectories as distributions induced by uncertainty ellipsoids to incorporate dynamics effects beyond action-space sampling, then enforces separation with Hellinger distance. For nonlinear or constrained dynamics the ellipsoid construction (typically via local linearization or moment propagation) can produce loose or biased support; nothing in the argument supplies an error bound showing that the resulting Hellinger metric still guarantees improved coverage or avoids new local minima. Experiments report aggregate speed-ups but do not isolate whether gains survive when the ellipsoid approximation is replaced by a higher-fidelity representation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes Uncertainty Guided Exploratory Trajectory Optimization (UGE-TO) for sampling-based model predictive control. Trajectories are represented as probability distributions induced by uncertainty ellipsoids that incorporate both action selection and system dynamics effects. Distributional separation is enforced via the Hellinger distance to promote better coverage of the configuration space and reduce convergence to local minima. The method is integrated into UGE-MPC, with experiments claiming 72.1% faster convergence in obstacle-free settings and 66% faster convergence plus 6.7% higher success rate in cluttered environments relative to the best baseline, under fixed sampling budgets. Validation includes simulations and real-world robot experiments, with code released.","tokens_in":1937,"tokens_out":669,"duration_ms":27782,"significance":"If the performance claims hold under rigorous scrutiny, the approach could meaningfully improve robustness of sampling-based MPC by explicitly accounting for dynamics-induced uncertainty in exploration, rather than relying solely on action-space sampling. The open-source code and real-world validation are positive contributions that support reproducibility. The quantitative speed-ups are notable, but their broader significance hinges on whether the ellipsoid approximation reliably yields superior coverage without new local minima or prohibitive overhead in nonlinear regimes.","major_comments":[{"comment":"§3.2 (Uncertainty Ellipsoid Construction): The claim that ellipsoid-induced distributions capture dynamics effects beyond action-space sampling and enable improved Hellinger-based separation rests on an unquantified approximation. For nonlinear or constrained dynamics, local linearization or moment propagation can yield loose or biased support; no error bound, sensitivity analysis, or comparison to Monte Carlo sampling is provided to show that the resulting Hellinger metric still guarantees better configuration-space coverage or avoids introducing new local minima. This is load-bearing for the central motivation and the reported gains.","section":"§3.2"},{"comment":"§4 (Experiments and Results): The reported metrics (72.1% faster convergence, 66% faster with 6.7% higher success rate) are aggregate figures without stated trial counts, standard deviations, or statistical significance tests. No ablation replaces the ellipsoid representation with a higher-fidelity distribution sampler to isolate whether gains survive when the approximation is relaxed, leaving the skeptic's concern about biased support unaddressed. This weakens verification of the performance claims under the fixed sampling budget.","section":"§4"}],"minor_comments":[{"comment":"Figure 3 and 4: Trajectory visualizations would benefit from overlaid uncertainty ellipsoids and explicit Hellinger distance annotations to directly illustrate the separation mechanism.","section":"Figures 3-4"},{"comment":"Notation in §3.1: Ensure the Hellinger distance formula and all parameters (e.g., covariance terms derived from ellipsoids) are defined before first use; a few symbols appear without prior introduction.","section":"§3.1"},{"comment":"The abstract states validation across 'a range of simulation scenarios,' yet §4 primarily details two environments; a brief summary table of additional scenarios would improve clarity.","section":"Abstract and §4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on our manuscript. We provide point-by-point responses to the major comments and indicate the revisions we will make.","responses":[{"response":"The ellipsoid approximation is indeed based on local linearization, which is a standard technique for uncertainty propagation in MPC to maintain computational efficiency. We do not claim theoretical guarantees on coverage or avoidance of local minima; the benefits are demonstrated empirically. To strengthen the paper, we will add a new subsection or appendix with a sensitivity analysis that compares the Hellinger distances computed from ellipsoids versus Monte Carlo samples of the nonlinear dynamics for representative trajectories. This will quantify the approximation error and its impact on sample diversity.","revision_made":"partial","referee_comment":"[§3.2] §3.2 (Uncertainty Ellipsoid Construction): The claim that ellipsoid-induced distributions capture dynamics effects beyond action-space sampling and enable improved Hellinger-based separation rests on an unquantified approximation. For nonlinear or constrained dynamics, local linearization or moment propagation can yield loose or biased support; no error bound, sensitivity analysis, or comparison to Monte Carlo sampling is provided to show that the resulting Hellinger metric still guarantees better configuration-space coverage or avoids introducing new local minima. This is load-bearing for the central motivation and the reported gains."},{"response":"We will revise the experimental section to include the number of trials (50 per setting), standard deviations for all reported metrics, and results of statistical significance tests. Furthermore, we will incorporate an ablation study that uses Monte Carlo sampling to generate the trajectory distributions instead of the ellipsoid approximation, allowing direct comparison of performance under the same sampling budget. This will help isolate the contribution of the approximation.","revision_made":"yes","referee_comment":"[§4] §4 (Experiments and Results): The reported metrics (72.1% faster convergence, 66% faster with 6.7% higher success rate) are aggregate figures without stated trial counts, standard deviations, or statistical significance tests. No ablation replaces the ellipsoid representation with a higher-fidelity distribution sampler to isolate whether gains survive when the approximation is relaxed, leaving the skeptic's concern about biased support unaddressed. This weakens verification of the performance claims under the fixed sampling budget."}],"tokens_in":1558,"tokens_out":491,"duration_ms":51133,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core move here is to turn uncertainty ellipsoids into trajectory distributions and then enforce separation between them with Hellinger distance instead of sampling only in action space. That produces the reported improvements in exploration and convergence speed for UGE-MPC under fixed budgets.","headline":"The paper adds a Hellinger-distance separation step on uncertainty-ellipsoid trajectory distributions to sampling-based MPC and reports clear speed and success gains in experiments, but the ellipsoid approximation for nonlinear dynamics is unproven.","tokens_in":2442,"tokens_out":140,"would_cite":false,"duration_ms":29249,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Representing trajectories as uncertainty-induced distributions enables sampling-based MPC to explore the configuration space more effectively and converge faster than action-space sampling.","keywords":["trajectory optimization","model predictive control","sampling-based planning","uncertainty ellipsoids","Hellinger distance","exploratory sampling","robotics control","local minima avoidance"],"falsifier":"A direct comparison run in a cluttered simulation environment where UGE-MPC fails to show at least 60 percent faster convergence or at least 5 percent higher success rate than the strongest baseline under identical sampling limits.","tokens_in":2716,"feed_emoji":"🤖","tokens_out":648,"duration_ms":41658,"temperature":0.7,"pith_summary":"The paper introduces Uncertainty Guided Exploratory Trajectory Optimization (UGE-TO), which models candidate trajectories as probability distributions shaped by uncertainty ellipsoids rather than sampling only in action space. These distributions incorporate the effects of system dynamics, and the method enforces separation between them using the Hellinger distance to promote diversity. When embedded in sampling-based model predictive control as UGE-MPC, the approach yields higher exploration, faster convergence to good solutions, and improved success rates under a fixed sampling budget. A sympathetic reader cares because standard sampling methods frequently trap optimization in local minima in cluttered or complex settings, and this change targets the root cause of poor initialization sensitivity without raising the sample count.","feed_headline":"Uncertainty ellipsoids speed MPC convergence by guiding diverse samples","feed_subtitle":"Modeling trajectories as separated distributions from uncertainty ellipsoids improves exploration and success rates under fixed sampling in ","key_machinery":"Trajectory probability distributions induced by uncertainty ellipsoids, with separation measured by Hellinger distance, that generate diverse samples for optimization.","core_discovery":"UGE-TO represents trajectories as probability distributions induced by uncertainty ellipsoids. This representation captures both dynamics and action effects, unlike pure action-space sampling. The algorithm then enforces distributional separation via the Hellinger distance to generate well-separated samples, achieving systematic coverage of the configuration space and greater robustness to local minima when used inside sampling-based MPC.","pith_inferences":["The same ellipsoid-distribution idea could be tested in other sampling planners such as RRT variants to see whether dynamics-aware separation helps beyond MPC.","Because the separation acts after dynamics propagation, the approach may scale better than action-space methods when state dimension grows.","Pairing the uncertainty ellipsoids with online learning of dynamics uncertainty could further tighten the distributions and reduce required samples.","The technique might support safer navigation in environments with moving obstacles by maintaining explicit coverage of reachable sets."],"forward_implications":["Higher trajectory diversity reduces trapping in local minima during optimization.","Faster convergence holds in both obstacle-free and cluttered settings under fixed sample budgets.","Higher task success rates appear when environments require large deviations from nominal paths.","The method remains practical for real-time control, as validated in both simulation and hardware experiments.","Systematic exploration improves robustness specifically where standard sampling-based MPC struggles most."],"fun_headline_variants":["Uncertainty ellipsoids represent trajectories as separated distributions in MPC","Separated uncertainty distributions guide exploratory sampling in MPC","UGE-TO uses Hellinger distance for distributional separation in MPC trajectories","Uncertainty ellipsoids capture dynamics and actions for MPC sampling"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Enforcing separation between uncertainty-ellipsoid distributions will produce better coverage of the configuration space than action sampling without introducing new local minima or prohibitive extra computation.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty ellipsoids represent trajectories as separated distributions in MPC","Separated uncertainty distributions guide exploratory sampling in MPC","UGE-TO uses Hellinger distance for distributional separation in MPC trajectories","Uncertainty ellipsoids capture dynamics and actions for MPC sampling"]},"model":"grok-4.3","cost_usd":0.013949,"raw_usage":{"total_tokens":5966,"prompt_tokens":719,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":139490500,"prompt_tokens_details":{"text_tokens":719,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":5183,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":719,"tokens_out":64,"duration_ms":61994,"temperature":1.0,"reasoning_tokens":5183,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-10T14:51:49.873259+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A direct comparison run in a cluttered simulation environment where UGE-MPC fails to show at least 60 percent faster convergence or at least 5 percent higher success rate than the strongest baseline under identical sampling limits.","supporting_citations":[],"review_version":1}