Uncertainty-aware autonomous navigation
Research archive · 04 selected works
Probabilistic robotics · perception-based safety · control barrier functions · uncertainty-aware world models
Occupancy-map-based safety with memory of out-of-view obstacles
OGM-CBF constructs a control barrier function from occupancy grid maps and signed distance fields. It supports arbitrary obstacle shapes, retains previously observed obstacles, and was evaluated in CARLA and on an industrial mobile robot.
Uncertainty-aware trajectory forecasting for off-road robots
ProTerrain models spatially correlated uncertainty in terrain parameters and propagates it through a differentiable physics engine. Structured convolutional operators enable efficient, high-resolution probabilistic trajectory predictions.
Perception-driven safety for unknown environments
V-CBF constructs control barrier functions directly from RGB-D observations, allowing safe control around previously unknown obstacles of arbitrary shape. The method uses image-to-image translation and was demonstrated with an autonomous car in CARLA.
Design and evaluation of cost maps for vision-based CBFs
This thesis investigates cost maps that transform segmented unsafe regions into control-barrier-compatible representations. The methods were evaluated in CARLA using ISO 22737-based and custom metrics, then deployed on an industrial mobile robot.
Experiences