MPPI Planning with Gaussian-Based Human Cost Function for Social Navigation

Chinmay Mundane

REVIEW 4 major objections 5 minor 20 references

MPPI Planning with Gaussian-Based Human Cost Function for Social Navigation

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that shaping the cost of an MPPI planner around predicted pedestrian motion—as speed-scaled anisotropic Gaussian fields—yields zero collisions across 300 randomized corridor scenarios while keeping planning real-time.

desk verdict PGIF is a clean, honest incremental MPPI cost term; the zero-collision result is real in simulation but the experiments don't isolate the anisotropic mechanism. read the letter →

arxiv 2608.08323 v2 pith:AINZSIAA submitted 2026-08-08 cs.RO cs.SYeess.SY

Chinmay Mundane This is my paper · ORCID

classification cs.ROcs.SYeess.SY

keywords MPPIsocialnavigationpedestrianpredictionanisotropicGaussianfieldcollisionavoidancemobilerobotcrowdmodelpredictivecontrol

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that a sampling-based motion planner can navigate crowded corridors safely if the cost function penalizes not just where pedestrians are now, but where they are predicted to be over the whole planning horizon. The proposed Predictive Gaussian Interaction Fields (PGIF) wrap each predicted pedestrian position in an anisotropic Gaussian cost that stretches forward along the person's motion and grows wider as speed increases, producing a motion cone of danger. In 300 randomized two-dimensional corridor trials at three crowd densities, the planner with this cost term suffered zero collisions at every density, while a standard MPPI planner that treats pedestrians as static points collided in up to 82% of runs. If the result holds, it would show that a closed-form, parallel cost term can give real-time anticipatory avoidance without changing the underlying MPPI machinery.

What carries the argument

The central object is the Predictive Gaussian Interaction Field (PGIF): a time-indexed anisotropic Gaussian repulsive cost attached to each pedestrian at each horizon step. Its defining identity is the spread rule $\sigma_{\parallel} = 1.2 + 0.5 s$ for the forward direction against pedestrian speed $s$, with a smaller fixed spread behind and laterally, together with the cost $C_{\text{human}} = \sum_{\tau} \sum_i \exp\left(-\tfrac12\left[(d_{\parallel}/\sigma_{\parallel})^2 + (d_{\perp}/\sigma_{\perp})^2\right]\right)$. This carries the argument because it converts predicted positions into a smooth, direction-aware danger landscape: faster pedestrians project longer danger cones ahead, and trajectories entering the cone receive exponentially higher cost, steering MPPI's weighted control update away from future collisions. Because it is a closed-form sum over $N$ pedestrians and $T$ horizon steps, it evaluates in $O(K \cdot T \cdot N)$ and parallelizes across the $K$ MPPI rollouts.

What would settle it

A concrete experiment: add zero-mean Gaussian noise with realistic tracker magnitude to the pedestrian positions and velocities fed to the PGIF cost at every planning step, rerun the same 300 scenarios, and count collisions; if the collision rate leaves zero, the demonstrated safety depends on perfect prediction rather than on the anisotropic field itself. A second test would replace the circular pedestrians with pedestrians who change direction mid-horizon, making the predictor wrong and exposing the cost field's robustness.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that predicting pedestrian motion and encoding it as speed-scaled anisotropic Gaussian repulsive fields in the MPPI cost yields zero collisions in all 300 tested scenarios while keeping planning real-time, at roughly 3 ms per step. The field is defined by decomposing the displacement from each predicted pedestrian position into a component parallel to motion and a perpendicular component, with a parallel standard deviation of 1.2 + 0.5 times the pedestrian's speed for the forward zone and 0.5 m behind, and a fixed 0.6 m lateral spread. The cost accumulates the Gaussian response over the full 40-step, 4-second horizon and over all pedestrians, so a rollout crossing a pedestrian's predicted path is penalized more than one passing behind. The paper is explicit that this is a soft cost, not a hard constraint, so it offers no formal safety guarantee; the zero-collision record is an empirical result in the simulated scenarios tested.

Load-bearing premise

The zero-collision result assumes the planner's prediction model is exactly the simulator's motion model—straight-line pedestrians follow constant velocity and circular pedestrians follow known circular kinematics, with current states read directly from the simulator—so under real perception noise or unexpected pedestrian behavior, the same cost field may not prevent collisions.

Editorial extensions

If this is right

Replacing static-point pedestrian obstacles with speed-scaled anisotropic fields removes collisions in every one of the 300 tested scenarios, dropping the collision rate from up to 82% to 0.
The predictive cost term is closed-form and parallelizable, so adding prediction to MPPI costs negligible computation: reported per-step planning times stay around 3 ms, meeting real-time constraints.
Safety comes at a price in dense crowds: the PGIF planner's success rate falls and its timeout rate rises to 59% at the hardest density, with longer average path lengths, showing that predicted risk can make the robot overly cautious.
Because PGIF is a single additive cost term, it can be inserted into any existing MPPI planner without changing the sampling procedure or control update.
The zero-collision claim is conditional on the evaluation setup, in which pedestrian states come directly from the simulator; real deployment would require perception and tracking to supply those states.
If this result transfers to physical robots, it gives a lightweight, learning-free route to anticipatory crowd navigation that improves on planners which only react to current pedestrian positions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension follows directly: because the PGIF cost is agnostic to the prediction source, replacing the constant-velocity and circular predictors with a learned goal-conditioned predictor would show whether the anisotropic encoding preserves its benefits under richer, more realistic motion models; the paper leaves this integration open.
We infer that the speed-scaled forward spread encodes an implicit time-to-arrival prior: a pedestrian moving at 1.2 m/s projects a danger cone roughly 2.2 m long, effectively reserving corridor space in proportion to how soon the person will occupy it. Tuning that scaling to prediction uncertainty, as the paper mentions as future work, would be a natural next experiment.
The high-density timeout behavior suggests that the fixed large human cost weight may be over-weighting predicted risk; reducing that weight adaptively when few low-cost rollouts exist—a direction the paper names—would likely trade a small collision risk for a much higher success rate, and this trade-off is directly testable in the same simulator.
A direct stress test would inject perception noise into the pedestrian states fed to the planner; if the zero-collision record degrades quickly with noise, the demonstrated safety depends on perfect prediction, not on the anisotropic cost field alone.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a clean, honest incremental contribution, and the zero-collision claim is likely real within the simulator—but the experiments don't isolate the proposed mechanism. I'd send it to review, expecting revisions.

What's actually new: a closed-form anisotropic Gaussian cost oriented along predicted pedestrian velocity, with forward spread growing linearly with speed. The math is straightforward, the code is public, and the write-up is refreshingly candid about being a soft cost with no formal guarantee. The comparison against vanilla MPPI shows a large effect, and the discussion of conservatism (timeout rising to 59%) is honest and useful.

Soft spots, in rough order of size. First, the evaluation leaks the future: circular pedestrians are predicted using the exact generative model parameters (orbit center, radius, angular velocity) the simulator itself uses. That's perfect prediction, not a fair test. Real perception would add noise, and the 0% number would not transfer. Second, the ablation is missing the key control: PGIF changes three things at once (predictive positions, anisotropy, and a huge weight w_human = 1e6). Without a predictive-isotropic baseline or a w_human sweep, you cannot attribute the collision reduction to the anisotropic motion cone. A predictive isotropic field at that weight would probably also avoid collisions, and in dense scenes it would also stall. The paper's own claim is that the speed-scaled forward spread is the meaningful part, but the data don't isolate it. Third, no error bars or statistical tests across the 100 runs per density—a minor fix, but needed given the success rates differ by ~20 points.

None of these are fatal. The equations are right, the method is simple and reproducible, and the limitations section says plainly that ground-truth states are used. The fix is to add a predictive-isotropic baseline, sweep w_human across at least a couple orders of magnitude, and run with noisy or learned predictions. Then the zero-collision claim would be properly supported.

Who it's for: people working on sampling-based MPC for social navigation, especially practitioners who want a cheap, parallel cost term. It's a useful reference, but I wouldn't treat it as a strong empirical demonstration. And yes, it deserves a serious referee—conditional accept, with the missing baseline as the main request.

Referee Report

4 major / 5 minor

Summary. The paper proposes Predictive Gaussian Interaction Fields (PGIF), a cost formulation for MPPI-based social navigation in which each pedestrian's predicted position over the planning horizon is encoded as an anisotropic Gaussian repulsive field aligned with the pedestrian's velocity. The forward spread of the kernel grows with pedestrian speed, creating a direction-dependent 'motion cone' cost that penalizes trajectories entering a pedestrian's future path more strongly than trajectories passing behind. The method is evaluated in a custom 2D corridor simulator with linear and circular pedestrian motion, at three crowd density levels, against a vanilla MPPI baseline that treats pedestrians as static point obstacles. The reported results show 0% collisions across all 300 trials for PGIF-MPPI versus up to 82% collisions for the baseline, with comparable per-step computation times, but also timeout rates up to 59% at high density. The paper claims the anisotropic, speed-dependent spread is the mechanism behind the safety improvement, and discusses conservatism and ground-truth-state limitations as future work.

Significance. If the central claim were established, PGIF would be a useful, lightweight addition to MPPI-based social navigation: the cost is closed-form, parallelizable across rollouts, independent of the choice of pedestrian predictor, and the paper provides an open-source implementation. The paper is also commendably transparent about its limitations, explicitly stating that PGIF is a soft cost with no formal safety guarantee (Section 3.5) and that the evaluation relies on ground-truth pedestrian states (Section 6). However, the empirical evidence as presented does not isolate the proposed anisotropic motion-cone mechanism, because the comparison changes prediction horizon, kernel shape, and cost weight simultaneously, and the planner is given the simulator's exact generative motion models. The significance of the work therefore depends on a strengthened evaluation that separates the effect of the kernel design from the effect of predictive information and extreme weight settings.

major comments (4)

[Section 3.6, Eq. (16); Table 1; Section 5] The reported 0% collision rate does not isolate the anisotropic PGIF mechanism. The PGIF-vs-vanilla comparison changes at least three things at once: (i) the cost uses future predicted positions instead of current positions, (ii) the kernel is anisotropic with sigma_parallel != sigma_perp (Eq. 13), and (iii) the human-cost weight is w_human = 1e6, six orders of magnitude above the other weights. Because no predictive-isotropic baseline and no w_human sweep are reported, the results are compatible with the weaker explanation that any strong repulsive cost applied to predicted positions eliminates collisions in this simulator. The authors should add an isotropic predictive Gaussian baseline and a sweep over w_human (including lower values) to determine whether the speed-scaled forward spread is actually the operative cause. As written, the experiments support only the claim that a predictive cost with very high weight avoids collisions in this simulator.
[Section 3.4, Eqs. (8)-(9); Section 4; Section 6] The planner's pedestrian prediction models are exactly the simulator's generative models: constant-velocity motion for linear pedestrians and exact circular kinematics with known orbit center, radius, and angular velocity for circular pedestrians. Section 4 states that pedestrian states are obtained directly from the simulator. The 0% collision result is therefore an oracle-prediction upper bound, not a property that can be expected to transfer to real perception and prediction pipelines. The paper acknowledges this in Section 6, but the abstract and Section 5 state '0% collision rate' without this caveat. The authors should either rephrase the central claim as an oracle-prediction result or add experiments with noisy/mismatched pedestrian predictions (e.g., added perception noise, wrong circular parameters, or a learned predictor) to quantify the sensitivity of the safety claim.
[Section 4; Table 1] The evaluation reports point estimates over 100 trials per condition with no variance, confidence intervals, or statistical tests. The headline claim is a collision rate of exactly zero, which is a point estimate that would be more informative with binomial confidence intervals; similarly, the 78% versus 93% success rates in the easy condition are not shown to be statistically distinguishable. The authors should report seeds, per-trial distributions, and confidence intervals or significance tests, and should consider adding a second predictive baseline such as DRAMPPI [17] to support the comparative claim.
[Section 5; Table 1] The framing of the central result is misleading in an important respect: at the hard density level, PGIF has a 59% timeout rate and only 41% success, so '0% collision rate' coexists with most episodes ending in the robot stalling before reaching the goal. The abstract's phrasing 'achieves a 0% collision rate at every density level' should be accompanied by the success/timeout rates for the same conditions, and the paper should discuss whether a plan that mostly times out in dense scenarios should be described as eliminating collisions, as opposed to avoiding collisions by inaction.

minor comments (5)

[Section 3.5, Eq. (13)] The units of the coefficients in Eq. (13) are implicit: sigma_parallel has units of meters, so 1.2 and 0.5 must carry units of meters and seconds, respectively, and sigma_perp = 0.6 is in meters. Please state these units explicitly.
[Section 5] The sentence 'The shorter paths is observed in the baseline at higher densities' contains a grammatical error and should be rephrased. In addition, the comparison 'the average path length increases from 14.43m to 18.75m' mixes the PGIF easy-condition mean (14.43) with the PGIF hard-condition mean (18.75); the relevant comparison for the high-density row is vanilla 6.66m versus PGIF 18.75m.
[Section 2.4] The text mentions Social LSTM, Social GAN, and Trajectron++ by name but does not give citations for them; please add the relevant references or remove the names.
[Figure 1] The caption says pedestrians are drawn as orange circles, but the figure appears in grayscale in the preprint; please ensure the color description matches the rendered figure or use a distinct marker style.
[Section 4] The baseline 'vanilla MPPI with static obstacle cost' is not precisely specified; please state whether the static cost uses the current pedestrian positions as point obstacles and what radius or kernel width is used, so that the comparison is reproducible.

Circularity Check

1 steps flagged · score 6.0 of 10

PGIF's zero-collision result is partly circular because the planner is given the simulator's own generative model as its pedestrian prediction, while the anisotropic kernel itself remains a non-circular design ansatz.

self definitional [Section 3.4 (Eq. 9); Section 4 (Experimental Setup); Section 6 (Discussion/Limitations)]
"For pedestrians following circular motion, we propagate the angular position as hτx,i = cx,i + ri cos(θi + ωiτ∆t), hτy,i = cy,i + ri sin(θi + ωiτ∆t). ... Pedestrian states are obtained directly from the simulator. ... The current evaluation relies on ground-truth pedestrian states provided by the simulator."

The planner's 'predicted' pedestrian positions are generated by the same kinematic equations and exact parameters that drive the simulator's own pedestrians, so the PGIF cost at each horizon step is evaluated on ground-truth future occupancy. The 0% collision rate is therefore an oracle result: by construction, the planner is told where pedestrians will be, and the cost term then penalizes exactly those positions. This does not test whether the anisotropic motion-cone encoding avoids collisions under imperfect prediction; the empirical safety claim reduces to feeding the simulator's generative model back into the cost. The paper's own Section 6 limitation confirms the input is ground truth, and the dominant human weight (w_human = 10^6, Eq.

full rationale

The derivation of the PGIF cost itself (Eqs. 13-14) is a transparent ansatz rather than a derivation from first principles; the 'motion cone' property is definitional in the kernel and is not circular by itself. I find no self-citation chain, no uniqueness theorem imported from the authors, and no fitted parameter renamed as a prediction: w_human = 10^6 is fixed across all scenarios and no tuning procedure is claimed, so the dominant safety weight is a design choice rather than a fitted input. The one significant circularity is in the evaluation: the 'pedestrian predictions' used by PGIF are exactly the simulator's generative models with ground-truth parameters (Eqs. 8-9; Section 6). Consequently, the headline 0% collision result is partly forced by construction, because the planner is given the true future occupancy at planning time and the human-cost term overwhelmingly dominates L(k) (Eq. 16). This reduces the strength of the central empirical claim with respect to real prediction error, although the anisotropic cost formulation itself retains independent content and the computational overhead claim is separately supported. I therefore set score 6 for partial circularity.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

PGIF is a cost formulation, not a new physical entity; no new force, particle, or dimension is introduced. The free parameters are hand-chosen cost shapes and weights, and the key domain assumption is that the planner has perfect knowledge of the simulator's pedestrian motion models.

free parameters (7)

sigma_parallel intercept = 1.2 m
Constant forward spread offset in Eq. (13); hand-chosen, no sensitivity analysis.
sigma_parallel speed slope = 0.5
Scales forward spread with pedestrian speed in Eq. (13); hand-chosen.
sigma_perp = 0.6 m
Lateral Gaussian width in Eq. (13); fixed and hand-chosen.
w_human = 1e6
Human cost weight in Eq. (16); extreme value that makes avoiding predicted fields dominate all other objectives and drives the conservative stalling behavior.
w_goal, w_term, w_path = 1000, 1000, 5000
Weights in Eq. (16); fixed across scenarios.
MPPI temperature lambda = 1000
Sampling temperature in Eq. (7); fixed.
scenario design parameters = 1/4/7 pedestrians; speeds 0.5-1.2 m/s; radii 1.2-2.0 m; angular rates 0.4-0.8 rad/s
Simulator sampling ranges define the three density levels; chosen by the author, not from a standard benchmark.

assumptions (3)

domain assumption Pedestrians move according to constant-velocity straight lines or fixed circular orbits throughout each episode.
Section 3.4 and Section 4 define the simulator's generative motion models; the planner's predictor uses the same models, so prediction error is zero by construction.
domain assumption The planner receives ground-truth pedestrian positions, velocities, and motion parameters.
Section 6 states 'Pedestrian states are obtained directly from the simulator'; Eq. (9) requires knowledge of orbit centers, radii, and angular velocities.
ad hoc to paper The anisotropic Gaussian cost (Eqs. 13-14) with chosen sigmas is an appropriate social-penalty model.
The cost shape is a design choice; the paper gives no ablation, comparison against isotropic predictive costs, or justification from human motion data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MPPI Planning with Gaussian-Based Human Cost Function for Social Navigation." pith.science (2026). https://pith.science/paper/AINZSIAA

@misc{pith2026260808323,
  author       = {Pith},
  title        = {Pith review of: MPPI Planning with Gaussian-Based Human Cost Function for Social Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AINZSIAA}},
  note         = {Machine review of arXiv:2608.08323}
}

read the original abstract

Safe robot navigation in crowded spaces requires planning that accounts for where people will be, not only where they are now. Model Predictive Path Integral (MPPI) control is an effective sampling-based planner, but many implementations encode humans as static point obstacles at their current positions, underestimating risk in dynamic scenes. We propose Predictive Gaussian Interaction Fields (PGIF), a spatiotemporal cost formulation that propagates pedestrian predictions forward over the full planning horizon and encodes them as anisotropic Gaussian repulsive fields aligned with each pedestrian's direction of motion. The forward spread of each field grows with the pedestrian's speed, creating a motion cone danger zone that penalises robot trajectories entering the pedestrian's path of travel more strongly than those approaching from behind. The formulation is closed-form and fully parallelisable across rollouts, adding no measurable computational overhead. Evaluated over 300 randomised crowd scenarios at three density levels, PGIF-MPPI achieves a 0% collision rate at every density level, compared with up to 82% for vanilla MPPI, while maintaining real-time planning performance.

Figures

Figures reproduced from arXiv: 2608.08323 by the authors.

**Figure 1.** Left: MPPI sample trajectories at a single planning step. The selected trajectory deviates above the direct path to avoid pedestrians (orange circles), each of which projects an anisotropic cost field elongated in their direction of travel. Right: The corresponding 2D anisotropic Gaussian cost field. High cost regions are elongated ahead of each pedestrian and narrower laterally and behind, reflecting the motion co… view at source ↗

**Figure 2.** Vanilla MPPI failure case. The robot selects a trajectory that is collision-free with respect to current pedestrian positions, but pedestrians move into its path during execution, leading to collision. trajectories that appear feasible at the current instant but intersect with pedestrian motion within the execution horizon. This contrast is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** PGIF-MPPI behaviour in a comparable scenario. The planner deviates early from the nominal path, avoiding regions that are predicted to become occupied, resulting in a longer but collision-free trajectory. 6 Discussion We employ two complementary prediction models: a constant-velocity model for pedestrians moving along straight trajectories, and exact circular kinematics for pedestrians following orbital motion. Thi… view at source ↗

Review history (3 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 3 linked inside Pith

[17]

Trevisan, E., Mustafa, K., Notten, G., Wang, X., Alonso-Mora, J.: Dynamic risk- aware mppi for mobile robots in crowds via efficient monte carlo approximations. pp. 313–320 (10 2025).https://doi.org/10.1109/IROS60139.2025.11246822

arXiv 2025
[1]

IEEE Transactions on Automatic Control62(8), 3861–3876 (2016)

Ames, A.D., Xu, X., Grizzle, J.W., Tabuada, P.: Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control62(8), 3861–3876 (2016)

2016
[2]

IEEE Robotics and Automation Letters11(1), 1–8 (2026).https://doi.org/10.1109/ LRA.2025.3630871

Belvedere, T., Ziegltrum, M., Turrisi, G., Modugno, V.: Feedback-mppi: Fast sampling-based mpc via rollout differentiation – adios low-level controllers. IEEE Robotics and Automation Letters11(1), 1–8 (2026).https://doi.org/10.1109/ LRA.2025.3630871

arXiv 2026
[3]

van den Berg, J., Lin, M., Manocha, D.: Reciprocal velocity obstacles for real- time multi-agent navigation. pp. 1928–1935 (05 2008).https://doi.org/10.1109/ ROBOT.2008.4543489

arXiv 1928
[4]

In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Cai, X., Everett, M., Fink, J., How, J.P.: Risk-Aware Off-Road Navigation via a Learned Speed Distribution Map. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2931–2937. IEEE (2022)

2022
[5]

In: 2019 In- ternational Conference on Robotics and Automation (ICRA)

Chen, C., Liu, Y., Kreiss, S., Alahi, A.: Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning. In: 2019 In- ternational Conference on Robotics and Automation (ICRA). p. 6015–6022. IEEE Press (2019).https://doi.org/10.1109/ICRA.2019.8794134,https://doi.org/ 10.1109/ICRA.2019.8794134

arXiv 2019
[6]

08862 PGIF MPPI for Social Navigation 13

Chen, Y.F., Everett, M., Liu, M., How, J.: Socially aware motion planning with deep reinforcement learning (03 2017).https://doi.org/10.48550/arXiv.1703. 08862 PGIF MPPI for Social Navigation 13

doi:10.48550/arxiv.1703 2017
[7]

In: Cavallo, F., Cabibihan, J.J., Fiorini, L., Sorrentino, A., He, H., Liu, X., Matsumoto, Y., Ge, S.S

De Benedictis, R., Beraldo, G., Cortellessa, G., Fracasso, F., Cesta, A.: A transformer-based approach for choosing actions in social robotics. In: Cavallo, F., Cabibihan, J.J., Fiorini, L., Sorrentino, A., He, H., Liu, X., Matsumoto, Y., Ge, S.S. (eds.) Social Robotics. pp. 198–207. Springer Nature Switzerland, Cham (2022)

2022

Show all 20 references

[8]

Han, T., Bao, Y., Mehta, B., Guo, G., Vishwakarma, A., Kang, E., Jung, S., Scalise, R., Zhou, J., Xu, B., Boots, B.: Model predictive adversarial imitation learning for planning from observation (2026),https://arxiv.org/abs/2507.21533

2026
[9]

Helbing, D., Molnár, P.: Social force model for pedestrian dynamics. Phys. Rev. E51, 4282–4286 (May 1995).https://doi.org/10.1103/PhysRevE.51.4282, https://link.aps.org/doi/10.1103/PhysRevE.51.4282

1995 doi
[10]

Kirby, R., Simmons, R., Forlizzi, J.: Companion: A constraint-optimizing method for person-acceptable navigation. pp. 607 – 612 (11 2009).https://doi.org/10. 1109/ROMAN.2009.5326271

2009
[11]

Robotics Auton

Kruse, T., Pandey, A.K., Alami, R., Kirsch, A.: Human-aware robot naviga- tion: A survey. Robotics Auton. Syst.61, 1726–1743 (2013),https://api. semanticscholar.org/CorpusID:5889262

2013
[12]

CoRR abs/1708.02596(2017),http://arxiv.org/abs/1708.02596

Nagabandi, A., Kahn, G., Fearing, R.S., Levine, S.: Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning. CoRR abs/1708.02596(2017),http://arxiv.org/abs/1708.02596

2017 arXiv
[13]

O’Callaghan, S.T., Ramos, F.T.: Gaussian process occupancy maps*. Int. J. Rob. Res.31(1), 42–62 (Jan 2012).https://doi.org/10.1177/0278364911421039, https://doi.org/10.1177/0278364911421039

2012 doi
[14]

Pellegrini, S., Ess, A., Van Gool, L.: You’ll never walk alone: Modeling social behavior for multi-target tracking. pp. 261–268 (09 2009).https://doi.org/10. 1109/ICCV.2009.5459260

2009
[15]

CoRRabs/1905.06113(2019), http://arxiv.org/abs/1905.06113

Rudenko, A., Palmieri, L., Herman, M., Kitani, K.M., Gavrila, D.M., Arras, K.O.: Human motion trajectory prediction: A survey. CoRRabs/1905.06113(2019), http://arxiv.org/abs/1905.06113

2019 arXiv
[16]

Trautman, P., Krause, A.: Unfreezing the robot: Navigation in dense, interacting crowds. pp. 797–803 (09 2010).https://doi.org/10.1109/IROS.2010.5654369

2010
[18]

Journal of Guidance, Control, and Dynamics 40(2), 344–357 (2017).https://doi.org/10.2514/1.G001921,https://doi.org/ 10.2514/1.G001921

Williams, G., Aldrich, A., Theodorou, E.A.: Model predictive path integral control: From theory to parallel computation. Journal of Guidance, Control, and Dynamics 40(2), 344–357 (2017).https://doi.org/10.2514/1.G001921,https://doi.org/ 10.2514/1.G001921

2017 doi
[19]

CoRRabs/1707.02342(2017),http://arxiv.org/abs/1707.02342

Williams, G., Drews, P., Goldfain, B., Rehg, J.M., Theodorou, E.A.: Information theoretic model predictive control: Theory and applications to autonomous driving. CoRRabs/1707.02342(2017),http://arxiv.org/abs/1707.02342

2017 arXiv
[20]

IEEE Robotics and Automation Letters4(2), 776–783 (2019)

Zhu, H., Alonso-Mora, J.: Chance-Constrained Collision Avoidance for MAVs in Dynamic Environments. IEEE Robotics and Automation Letters4(2), 776–783 (2019)

2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [17]

Trevisan, E., Mustafa, K., Notten, G., Wang, X., Alonso-Mora, J.: Dynamic risk- aware mppi for mobile robots in crowds via efficient monte carlo approximations. pp. 313–320 (10 2025).https://doi.org/10.1109/IROS60139.2025.11246822

arXiv 2025

[2] [1]

IEEE Transactions on Automatic Control62(8), 3861–3876 (2016)

Ames, A.D., Xu, X., Grizzle, J.W., Tabuada, P.: Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control62(8), 3861–3876 (2016)

2016

[3] [2]

IEEE Robotics and Automation Letters11(1), 1–8 (2026).https://doi.org/10.1109/ LRA.2025.3630871

Belvedere, T., Ziegltrum, M., Turrisi, G., Modugno, V.: Feedback-mppi: Fast sampling-based mpc via rollout differentiation – adios low-level controllers. IEEE Robotics and Automation Letters11(1), 1–8 (2026).https://doi.org/10.1109/ LRA.2025.3630871

arXiv 2026

[4] [3]

van den Berg, J., Lin, M., Manocha, D.: Reciprocal velocity obstacles for real- time multi-agent navigation. pp. 1928–1935 (05 2008).https://doi.org/10.1109/ ROBOT.2008.4543489

arXiv 1928

[5] [4]

In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Cai, X., Everett, M., Fink, J., How, J.P.: Risk-Aware Off-Road Navigation via a Learned Speed Distribution Map. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2931–2937. IEEE (2022)

2022

[6] [5]

In: 2019 In- ternational Conference on Robotics and Automation (ICRA)

Chen, C., Liu, Y., Kreiss, S., Alahi, A.: Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning. In: 2019 In- ternational Conference on Robotics and Automation (ICRA). p. 6015–6022. IEEE Press (2019).https://doi.org/10.1109/ICRA.2019.8794134,https://doi.org/ 10.1109/ICRA.2019.8794134

arXiv 2019

[7] [6]

08862 PGIF MPPI for Social Navigation 13

Chen, Y.F., Everett, M., Liu, M., How, J.: Socially aware motion planning with deep reinforcement learning (03 2017).https://doi.org/10.48550/arXiv.1703. 08862 PGIF MPPI for Social Navigation 13

doi:10.48550/arxiv.1703 2017

[8] [7]

In: Cavallo, F., Cabibihan, J.J., Fiorini, L., Sorrentino, A., He, H., Liu, X., Matsumoto, Y., Ge, S.S

De Benedictis, R., Beraldo, G., Cortellessa, G., Fracasso, F., Cesta, A.: A transformer-based approach for choosing actions in social robotics. In: Cavallo, F., Cabibihan, J.J., Fiorini, L., Sorrentino, A., He, H., Liu, X., Matsumoto, Y., Ge, S.S. (eds.) Social Robotics. pp. 198–207. Springer Nature Switzerland, Cham (2022)

2022