Dog Hit by a Delivery Robot: 3D Forensic Analysis of a Depth Error

A last-mile delivery robot ran over a dog on a wet sidewalk. The incident, captured by the vehicle's own cameras, has undergone a 3D expert analysis to clarify the cause. The main hypothesis points to a depth error induced by the specular reflection of a puddle of water, which would have deceived the stereo vision system.

Geometric reconstruction with OpenCV 3D and simulation in Gazebo 🛠️

The forensic workflow begins with extracting the stereo frames from the robot. In Python, using OpenCV 3D, the disparity map is calculated to generate a point cloud of the scene. When analyzing the puddle area, it is observed that the dog's reflection generated a false correspondence point, shifting the depth estimation several meters backward. With this data, the geometry is imported into Gazebo to reproduce the robot's trajectory and the exact moment of impact. The simulation confirms that, when crossing the puddle, the obstacle avoidance system interpreted that the animal was farther away, executing a late braking maneuver. Finally, in Blender, the scene is rendered with the visual error vectors to document the failure.

Lessons for delivery robotics safety ⚠️

This case demonstrates that reflective surfaces are a critical blind spot for stereo vision systems based on pixel correlation. The technical solution involves fusing the camera with active depth sensors, such as LIDAR, which are not affected by specular reflections. Furthermore, 3D expert analysis is consolidated as an indispensable tool for auditing accidents and improving perception algorithms in real urban environments.

Could the collision have been avoided if the robot's depth sensor had correctly interpreted the reflective texture of the wet ground? 🤔

(PS: In scene analysis, every scale witness is a small anonymous hero.)