Risk-Aware Model Predictive Control for Unmanned Ground Vehicles

Frank Gonzalez1, Varun Murali1, Fernando Cladera1, Camillo J. Taylor1, Vijay Kumar1,
1University of Pennsylvania,
Paper arXiv Video Code Data

Abstract

UGVs have been successfully deployed in both research and industrial applications in structured settings but are increasingly deployed in unstructured off-road settings. In this context, motion planning in the presence of risks is crucial. Traditionally, mobile robotics has focused on positive obstacles, stationary or moving, as the main source of risks. In this work, we focus on explicitly quantifying and avoiding negative obstacles in addition to positive ones. Negative obstacle detection in UGVs proves to be a key problem for autonomous navigation, as negative obstacles do not register on typical navigation sensors. Traditional obstacle detection techniques fail to identify these obstacles as dangerous, and therefore fail to incorporate them when generating obstacle maps. In this work, we propose a risk-aware planner that is able to combine different sources of risk. We focus on the problem of negative obstacle detection and outline steps for incorporating other risks into our planner. Our results show that negative obstacles are detected with mean Intersection over Union, Precision, Recall, and F1 scores of 0.54, 0.64, 0.77, and 0.69, enabling risk-avoidant navigation around ditches, curbs, and steep drop-offs. The system demonstrates adaptive path planning that circumvents high-risk areas, though performance is limited by occlusion effects and terrain sensitivity.

Full Video

Large Ditch Detection

Real-time detection of a large ditch that poses a navigation hazard. The system identifies the steep negative obstacle and generates risk maps (darker regions indicate higher risk) to enable safe path planning.

Curb Detection

Detection of sidewalk curbs that can trap the UGV despite their small size. The system successfully identifies these abrupt recessions in the terrain, though detection may be partial due to positive obstacle occlusions.

Pipeline Experiments

We conducted field experiments at four distinct sites around a Semi-Urban Office Park to evaluate the robustness of our risk-aware planner:

  • Large Ditch: A steep slope presenting a navigation hazard for the UGV.
  • Small Ditch: A traversable but visually salient depression.
  • Long Hill: An extended hillside with boundary drop-offs for path tracing evaluation.
  • Sharp Hill: A hill with sharp drop-offs testing high-risk terrain identification.

System Overview

System Overview

Autonomy architecture highlighting the overall flow of data between the subsystems. LiDAR points are ingested by DLIO to produce an estimate of the robot's pose. The point cloud is both fed into the negative obstacle detector and processed by groundgrid to segment non-ground points for the positive obstacle detector. Both detectors produces occupancy grids, which are combined to produce a risk map that is used by A* to get a reference path. This reference path is optimized to be dynamically feasible and sent to the low-level controller.

BibTeX