Interpretable Collision Avoidance and Feasibility Guarantees through an Uncertainty Aware Pipeline

Doris Xu; Jin Hyun Park; Guadalupe Alvarado Fernandez; Adhyan Prasad; Mark Campbell

cs.ROOct 23, 2026

Interpretable Collision Avoidance and Feasibility Guarantees through an Uncertainty Aware Pipeline

Authors: Doris Xu, Jin Hyun Park, Guadalupe Alvarado Fernandez, Adhyan Prasad, Mark Campbell

Organizations: The authors are with the Sibley School of Mechanical and Aerospace Engineering at Cornell University, Ithaca, NY, 14850.

Abstract

Autonomous vehicles (AVs) operating in dynamic, challenging conditions such as intersections and roundabouts must guarantee collision avoidance under uncertainty while maintaining real-time feasibility. Most existing planners struggle to simultaneously guarantee both requirements. Chance-constrained formulations embed safety constraints directly in the optimization but become infeasible as uncertainty compounds; deterministic planners provide feasible solutions but fail to address the uncertainty that causes collisions; end-to-end learned policies provide neither guarantee in an interpretable form. We present a modular perception-prediction-planning pipeline that closes this gap by propagating uncertainty as structured, interpretable quantities across module boundaries. Rather than embedding probabilistic risk directly in the optimization, our pipeline resolves collision probabilities upstream and constructs convex safe sets within which the planner provides a guaranteed feasible solution at every cycle -- decoupling the safety guarantee from solver feasibility. Across Monte Carlo trials in a four-way unsignalized intersection, our pipeline achieves 0% collisions and 0% solver infeasibility at 12.7ms per cycle, compared to 36% and 10% collision rates for deterministic and chance-constrained baselines respectively. The ablation study confirms the necessity of each uncertainty source and the hardware deployment on Clearpath Jackal UGVs validates the pipeline's real-time viability on physical platforms.

Figures & tables

Fig. 1: Multiple uncertainty sources in a two-vehicle intersection encounter. The agent’s discrete intent is unknown, its forecasted trajectories accumulate covariance over the prediction horizon, and both vehicles carry localization uncertainty. Each source independently degrades collision avoidance if neglected (Table II), motivating a pipeline that preserves all of them as structured quantities across module boundaries.

Fig. 2: Pipeline Architecture Overview: The IMM estimates the agent’s intent probabilities and states. The forecaster produces the forecasted trajectories. The collision detector combines these forecasted trajectories with the ego localizer’s state estimates to compute per-intent collision probabilities. The decision maker converts these probabilities into a convex safe set, within which the flat MPC plans and the CBF safety filter enforces forward invariance.

Fig. 3: Pipeline snapshot for a two-vehicle intersection encounter. (a) The IMM produces per-intent mode probabilities

\mu^{j}

; the forecaster propagates each into a trajectory (blue) with compounded covariance ellipses. (b) The collision detector finds the closest OBB approach

n^{*}

(blue markers) along each forecast. (c) Combined covariance

\Sigma

is projected onto the separation direction to yield

P_{\text{col}}^{j}

analytically. (d) The decision maker sweeps the agent’s OBB along the most likely forecast to form the unsafe region

\mathcal{U}

and constructs the convex safe set

\mathcal{S}

and determines the intermediate goal pose (shown as a red star) for the downstream planner and safety filter.

Planner	Scenario	$N$	Coll.%	Succ.%	OBB Dist (m)	Goal Time (s)	Norm. Time	Infeas.%	QP (ms)	Cycle (ms)
Ours	Keep Straight	35	0.0	100.0	$0.18\pm 0.03$	$13.07\pm 0.73$	$2.09\pm 0.12$	0.0	$4.7\pm 2.4$	$12.4\pm 2.5$
Ours	Left Turn	37	0.0	100.0	$0.26\pm 0.03$	$12.96\pm 0.94$	$2.07\pm 0.15$	0.0	$5.4\pm 3.4$	$14.2\pm 3.4$
Ours	Right Turn	28	0.0	100.0	$0.30\pm 0.01$	$9.16\pm 0.55$	$1.47\pm 0.09$	0.0	$5.2\pm 3.0$	$11.0\pm 3.0$
Ours	Overall	100	0.0	100.0	$0.25\pm 0.06$	$11.94\pm 1.90$	$1.91\pm 0.30$	0.0	$5.1\pm 2.9$	$12.7\pm 3.3$
MPC	Keep Straight	35	2.9	97.1	$0.08\pm 0.03$	$7.70\pm 0.71$	$1.23\pm 0.11$	0.0	$18.4\pm 3.2$	$18.4\pm 3.2$
MPC	Left Turn	37	94.6	5.4	$0.10\pm 0.00$	$11.20\pm 0.28$	$1.79\pm 0.05$	0.0	$23.5\pm 3.4$	$23.5\pm 3.4$

TABLE I: Monte Carlo Analysis Results

Planner	Scenario	$N$	Coll.%	Succ.%	OBB Dist (m)	Goal Time (s)	Norm. Time	QP (ms)	Cycle (ms)
Full	Keep Straight	35	0.0	100.0	$0.18\pm 0.03$	$13.07\pm 0.73$	$2.09\pm 0.12$	$4.7\pm 2.4$	$12.4\pm 2.5$
Full	Left Turn	37	0.0	100.0	$0.26\pm 0.03$	$12.96\pm 0.94$	$2.07\pm 0.15$	$5.4\pm 3.4$	$14.2\pm 3.4$
Full	Right Turn	28	0.0	100.0	$0.30\pm 0.01$	$9.16\pm 0.55$	$1.47\pm 0.09$	$5.2\pm 3.0$	$11.0\pm 3.0$
Full	Overall	100	0.0	100.0	$0.25\pm 0.06$	$11.94\pm 1.90$	$1.91\pm 0.30$	$5.1\pm 2.9$	$12.7\pm 3.3$
No Intent	Keep Straight	35	45.7	54.3	$0.19\pm 0.04$	$12.87\pm 0.51$	$2.06\pm 0.08$	$3.0\pm 0.1$	$8.0\pm 0.3$
No Intent	Left Turn	37	35.1	64.9	$0.23\pm 0.08$	$12.71\pm 1.36$	$2.03\pm 0.22$	$2.9\pm 0.1$	$8.7\pm 0.4$

TABLE II: Ablation Study Results

Fig. 4: Snapshots of three experiment scenarios at three different times: $t_{0}$ is the start time, $t_{1}$ is the time of highest collision probability, and $t_{2}$ is the time when the unsafe set no longer intersects with the lane. The forecasted trajectories are shown in blue (lines and covariance ellipsoids). The ego’s planned trajectory, goal point, and safe set are shown in yellow (as line, dot, and rectangle, respectively)

Fig. 5: The IMM estimator outputs oscillating intent probabilities as the agent follows a zig-zag trajectory not in $\mathcal{M}$ . Five timesteps are shown: $t_{1}$ —agent turns right; $t_{2}$ —switches to left; $t_{3}$ —switches to straight; $t_{4}$ —commits to left; $t_{5}$ —ego deems it safe to proceed.

Safe Motion Planning

Safe motion planning develops methods for robots and vehicles to reach goals while avoiding collisions and respecting dynamic and uncertainty constraints.

See all 214 papers

The Impact of Class Uncertainty Propagation in Perception-Based Motion Planning
Feb 17, 2026Jibran Iqbal Shah, Andrei Ivanovic, Kelly Zhu +4Autonomous Driving Planning Motion Planning
Department of Computer Science, University of Toronto. · Toyota Research Institute.
Exact, Efficient, and Safe Occlusion-Aware Planning Using AH-Polyhedrons
Jun 13, 2026Long Kiu Chung, David Isele, Toktam Mohammadnejad +4Reachability Analysis Autonomous Parking
Honda Research Institute (HRI), Mountain View, CA. · Georgia Institute of Technology, Atlanta, GA.
U-OBCA: Uncertainty-Aware Optimization-Based Collision Avoidance via Wasserstein Distributionally Robust Chance Constraints
Mar 5, 2026Zehao Wang, Yuxuan Tang, Han Zhang +2Chance-Constrained Motion Planning Collision Avoidance
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University and Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai 200240, China. · Key Laboratory of System Control and Information Processing, Ministry of Education of China, and Shanghai Engineering Research Center of Intelligent Control and Management, Shanghai 200240, China.

Feb 17, 2026J/K move · Enter open · S save

The Impact of Class Uncertainty Propagation in Perception-Based Motion Planning

Jibran Iqbal Shah, Andrei Ivanovic, Kelly Zhu, Masha Itkina+3

Department of Computer Science, University of Toronto. · Toyota Research Institute.

Autonomous Driving Planning Motion Planning Uncertainty Propagation Uncertainty Calibration

Open paper AskPDF arXiv

Autonomous vehicles (AVs) are being increasingly deployed in urban environments. In order to operate safely and reliably, AVs need to account for the inherent uncertainty associated with perceiving the world through sensor data and incorporate that into their decision-making process. Uncertainty-aware planners have recently been developed to account for upstream perception and prediction uncertainty. However, such planners may be sensitive to prediction uncertainty miscalibration, the magnitude of which has not yet been characterized. Towards this end, we perform a detailed analysis on the impact that perceptual uncertainty propagation and calibration has on perception-based motion planning. We do so by comparing two novel prediction-planning pipelines with varying levels of uncertainty propagation on the recently-released nuPlan planning benchmark. We study the impact of upstream uncertainty calibration using closed-loop evaluation on the nuPlan challenge scenarios. We find that the method incorporating upstream uncertainty propagation demonstrates superior generalization to complex closed-loop scenarios.

Abstract

Safe Motion Planning

Explore similar work

The Impact of Class Uncertainty Propagation in Perception-Based Motion Planning

Exact, Efficient, and Safe Occlusion-Aware Planning Using AH-Polyhedrons

U-OBCA: Uncertainty-Aware Optimization-Based Collision Avoidance via Wasserstein Distributionally Robust Chance Constraints