Risk-Sensitive Crowd Navigation with Adaptive Ellipsoidal Conformal Prediction

Ruihan A. Li; Ziyao Guo; Yingying Li

cs.ROOct 5, 2026

Risk-Sensitive Crowd Navigation with Adaptive Ellipsoidal Conformal Prediction

Authors: Ruihan A. Li, Ziyao Guo, Yingying Li

Organizations: Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, IL

Abstract

Safe crowd navigation under distribution shift requires uncertainty representations that capture structured human-motion prediction errors and safety objectives that account for rare but consequential failures. Existing uncertainty-aware methods typically represent prediction errors using isotropic regions, which can be either overly conservative or poorly aligned with directional motion uncertainty. We introduce a risk-aware navigation framework that uses anisotropic conformal ellipsoids to translate structured prediction uncertainty into an episode-level conditional value-at-risk (CVaR) signal that regulates the Lagrangian safety penalty. In particular, adaptive ellipsoidal conformal prediction (AECP) captures directional prediction errors and adaptively calibrates uncertainty regions under distribution shift, while the resulting CVaR-regulated navigation policy is optimized using Lagrangian proximal policy optimization. We evaluate our proposed method under both in-distribution settings and out-of-distribution (OOD) settings involving shifts in pedestrian motion patterns. Compared with state-of-the-art baselines, our method maintains competitive in-distribution performance while improving success rates by 5.68-7.44 percentage points and reducing collision rates by 5.44-6.64 percentage points across OOD settings. We further deploy the trained policy without fine-tuning on a physical robot with onboard perception and CPU-only inference, showing that the full pipeline is feasible in physical crowd navigation.

Figures & tables

Fig. 1 : Prediction errors from Gumbel Social Transformer two-step trajectory predictions for pedestrian 1 in the OOD crossing-flow scenario (see Section IV ). Grey points denote samples of prediction errors/residuals. The blue dashed contour illustrates a circular uncertainty region, while the red contour illustrates an ellipsoidal uncertainty region. The residuals exhibit clear directional anisotropy: a circular region must expand uniformly to cover errors along the dominant direction, whereas an ellipsoidal region can adapt both orientation and relative axis lengths.

Scenario	Method	SR $\uparrow$	CR $\downarrow$	TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	ITR $\downarrow$	SD $\uparrow$
In-distribution	SF	$11.68{\scriptstyle\pm 3.22}\%$	$21.28{\scriptstyle\pm 0.33}\%$	$67.04{\scriptstyle\pm 3.51}\%$	$30.72{\scriptstyle\pm 2.58}$	$27.17{\scriptstyle\pm 0.16}$	$3.82{\scriptstyle\pm 0.08}\%$	$0.432{\scriptstyle\pm 0.002}$
	ORCA	$64.48{\scriptstyle\pm 3.14}\%$	$31.28{\scriptstyle\pm 1.11}\%$	$4.24{\scriptstyle\pm 2.20}\%$	$24.07{\scriptstyle\pm 1.15}$	$13.14{\scriptstyle\pm 0.60}$	$1.49{\scriptstyle\pm 0.14}\%$	$0.490{\scriptstyle\pm 0.005}$
	CrowdNav++	$87.60{\scriptstyle\pm 3.32}\%$	$12.40{\scriptstyle\pm 3.32}\%$	$0.00{\scriptstyle\pm 0.00}\%$	$14.03{\scriptstyle\pm 0.33}$	$13.12{\scriptstyle\pm 0.32}$	$7.65{\scriptstyle\pm 0.12}\%$	$0.417{\scriptstyle\pm 0.009}$
	GenSafeNav	$97.28{\scriptstyle\pm 0.44}\%$	$2.72{\scriptstyle\pm 0.44}\%$	$0.00{\scriptstyle\pm 0.00}\%$	$17.51{\scriptstyle\pm 0.51}$	$17.11{\scriptstyle\pm 0.52}$	$3.07{\scriptstyle\pm 0.10}\%$	$0.4467{\scriptstyle\pm 0.0096}$
	Ours (AECP)	$97.76{\scriptstyle\pm 0.36}\%$	$2.24{\scriptstyle\pm 0.36}\%$	$0.00{\scriptstyle\pm 0.00}\%$	$16.98{\scriptstyle\pm 0.41}$	$16.66{\scriptstyle\pm 0.42}$	$4.29{\scriptstyle\pm 0.24}\%$	$0.419{\scriptstyle\pm 0.004}$
OOD-Crossing Flow	SF	$1.84{\scriptstyle\pm 0.54}\%$	$16.80{\scriptstyle\pm 0.40}\%$	$81.36{\scriptstyle\pm 0.36}\%$	$32.98{\scriptstyle\pm 1.43}$	$19.86{\scriptstyle\pm 0.49}$	$4.25{\scriptstyle\pm 0.71}\%$	$0.354{\scriptstyle\pm 0.003}$

TABLE I : Main test results. Results are reported as mean

\pm

standard deviation. Lower is better for CR, TR, NT, PL, and ITR; higher is better for SR and SD. ITR and SD reflect proximity-comfort tradeoffs.

Scenario	Method	SR $\uparrow$	CR $\downarrow$	TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	ITR $\downarrow$	SD $\uparrow$
OOD-Crossing Flow	Ours (Circular)	$72.32{\scriptstyle\pm 1.53}\%$	$25.52{\scriptstyle\pm 1.75}\%$	$2.16{\scriptstyle\pm 0.22}\%$	$18.06{\scriptstyle\pm 0.86}$	$15.86{\scriptstyle\pm 0.25}$	$7.86{\scriptstyle\pm 0.19}\%$	$0.3669{\scriptstyle\pm 0.0029}$
OOD-Crossing Flow	Ours (AECP)	$77.68{\scriptstyle\pm 1.61}\%$	$21.36{\scriptstyle\pm 1.71}\%$	$0.96{\scriptstyle\pm 0.22}\%$	$17.38{\scriptstyle\pm 0.49}$	$15.31{\scriptstyle\pm 0.59}$	$7.87{\scriptstyle\pm 0.25}\%$	$0.368{\scriptstyle\pm 0.003}$
OOD-Merging Flow	Ours (Circular)	$80.64{\scriptstyle\pm 2.71}\%$	$15.20{\scriptstyle\pm 2.93}\%$	$4.16{\scriptstyle\pm 0.22}\%$	$28.34{\scriptstyle\pm 0.73}$	$26.44{\scriptstyle\pm 0.97}$	$5.04{\scriptstyle\pm 0.28}\%$	$0.3871{\scriptstyle\pm 0.0037}$
OOD-Merging Flow	Ours (AECP)	$84.80{\scriptstyle\pm 2.76}\%$	$13.28{\scriptstyle\pm 0.87}\%$	$1.92{\scriptstyle\pm 1.95}\%$	$27.15{\scriptstyle\pm 0.78}$	$25.08{\scriptstyle\pm 0.27}$	$6.14{\scriptstyle\pm 0.20}\%$	$0.388{\scriptstyle\pm 0.002}$
OOD-Dynamic Velocity	Ours (Circular)	$74.48{\scriptstyle\pm 3.06}\%$	$25.12{\scriptstyle\pm 3.06}\%$	$0.40{\scriptstyle\pm 0.00}\%$	$17.00{\scriptstyle\pm 0.31}$	$14.87{\scriptstyle\pm 0.07}$	$5.06{\scriptstyle\pm 0.37}\%$	$0.4063{\scriptstyle\pm 0.0050}$
	Ours (AECP)	$75.44{\scriptstyle\pm 2.63}\%$	$24.32{\scriptstyle\pm 2.42}\%$	$0.24{\scriptstyle\pm 0.22}\%$	$16.95{\scriptstyle\pm 0.33}$	$14.83{\scriptstyle\pm 0.21}$	$5.59{\scriptstyle\pm 0.43}\%$	$0.406{\scriptstyle\pm 0.002}$

TABLE II : Comparison between AECP and adaptive circular conformal prediction under OOD conditions. Both variants use DtACI-based online scale adaptation and the same CVaR-regulated PPO-Lagrangian training pipeline and differ only in uncertainty-set geometry. Results are reported as mean

\pm

standard deviation. Lower is better for CR, TR, NT, PL, and ITR; higher is better for SR and SD.

Scene	$\mathbf{k}$	$\mathbf{t}$	Circ. Cov.	Ellip. Cov.	Circ. Area	Ellip. Area
Crossing Flow	1	10	88.03%	88.40 %	0.102	0.088
Crossing Flow	1	40	52.28%	74.03 %	1.071	0.555
Crossing Flow	1	80	91.38 %	89.66%	2.064	0.453
Crossing Flow	2	10	89.51%	94.04 %	0.358	0.276
Crossing Flow	2	40	67.75%	78.28 %	0.985	0.569
Crossing Flow	2	80	89.66%	96.55 %	1.371	0.447

TABLE III : Circular and ellipsoidal coverage statistics at different prediction horizons

1\leq k\leq 2

and at evaluation times

t=10,40,80

for demonstration.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Humans / arena size	$20$ / $12\times 12$ m
Simulation step / time limit	$0.25$ s / $50$ s
DtACI target miscoverage	$\alpha=0.10$
Prediction horizon / Considered Step	$5$ / $2$
Deployed DtACI radius	$\bar{\tau}\in[0,1]$
CVaR confidence level	$\beta=0.95$

Appendix

TABLE IV : Implementation and training parameters.

Explore similar work

DiffCVaR: Reinforcement Learning for Risk Adaptation via Differentiable CVaR Barrier Functions
May 20, 2026Xinyi Wang, Taekyung Kim, Bardh Hoxha +2Obstacle Avoidance Offline Reinforcement Learning
Department of Robotics, University of Michigan, Ann Arbor, MI, 48109, USA · Toyota Motor North America, Research & Development, Ann Arbor, MI, 48105, USA · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA
CoCoNav: Conformal Control for Safe Robot Navigation in Crowds
Aug 7, 2026Cheng Guo, Mingzhe Ni, Zheng Liang +5Robot Navigation Model Predictive Control
Department of Computer Science, The University of Manchester, Manchester, UK. · Human-Robot Interfaces and Interaction Laboratory, Italian Institute of Technology, Genoa, Italy. · Genisom AI, Shanghai, China. +3
Differentiable Optimization Layered Safety-Critical Control for Risk-Aware Navigation via Conformal Prediction
May 5, 2026Jinyang Dong, Shizhen Wu, Yongchun FangObstacle Avoidance Robust Trajectory
Institute of Robotics and Automatic Information Systems, College of Artificial Intelligence, Nankai University, Tianjin 300071, China · Academy for Advanced Interdisciplinary Studies, Nankai University, Tianjin 300071, China · Institute of Intelligence Technology and Robotic Systems, Shenzhen Research Institute of Nankai University, Shenzhen 518083, China

May 20, 2026, cs.ROJ/K move · Enter open · S save

DiffCVaR: Reinforcement Learning for Risk Adaptation via Differentiable CVaR Barrier Functions

Xinyi Wang, Taekyung Kim, Bardh Hoxha, Georgios Fainekos+1

Department of Robotics, University of Michigan, Ann Arbor, MI, 48109, USA · Toyota Motor North America, Research & Development, Ann Arbor, MI, 48105, USA · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA

Obstacle Avoidance Offline Reinforcement Learning Barriers Gaussian Mixture Models

Open paper AskPDF arXiv

Planning through crowded environments under uncertain obstacle motions remains difficult, as stochastic interactions often induce overly conservative behavior or reduced efficiency. To address this challenge, we propose an end-to-end risk adaptation framework for crowd navigation under obstacle-motion uncertainty modeled by a Gaussian mixture model. The framework combines reinforcement learning~(RL) with a differentiable quadratic-program safety layer based on Conditional Value-at-Risk~(CVaR) barrier functions, jointly learning nominal control input, risk level, and safety margin and enforcing explicit probabilistic safety constraints. This design enables context-aware adaptation, promoting efficient behavior while invoking caution only when necessary. We conduct extensive evaluations in dynamic, uncertain, and crowded environments across varying obstacle densities and robot models, and further assess generalization under three out-of-distribution cases. Comparisons across optimization-based, RL-based, and integrated RL and optimization methods are provided, and the proposed method is shown to deliver the strongest overall performance in safety, efficiency, and generalization under uncertainty.

Abstract

Explore similar work

DiffCVaR: Reinforcement Learning for Risk Adaptation via Differentiable CVaR Barrier Functions

CoCoNav: Conformal Control for Safe Robot Navigation in Crowds

Differentiable Optimization Layered Safety-Critical Control for Risk-Aware Navigation via Conformal Prediction