Autonomous Road Maintenance
Being part of the development of a product is always exciting. But when it is a self-driving robot that solves real-world problems, it is even more fun.
Detect defects in the road surface
Robotiz3d is building an autonomous road maintenance system that can identify and repair cracks and potholes. The system uses profilometry to create a 3D point cloud of the road surface that it uses to identify defects and then directly fix these in one working process.
Robotiz3D and Desupervised have got a RIMA grant to bring this to life. At Desupervised we’re focusing on analyzing the 3D point clouds to identify the various defects, using a combination of AI modeling techniques.
Since we got involved in the project at a very early stage, before the sensor equipment was actually ready, we started our work with synthetic data where we evaluated several point cloud model architectures for their suitability. We first landed on the PointNet++ architecture before finally transitioning to an architecture named RandLA-Net.
Since we got involved in the project at a very early stage, before the sensor equipment was actually ready, we started our work with synthetic data where we evaluated several point cloud model architectures for their suitability. We first landed on the PointNet++ architecture before finally transitioning to an architecture named RandLA-Net.
A typical point cloud viewed from the side where the y axis has been amplified for clarity, with the actual data on the left and the prediction to the right.
Having reached the desired model predictive performance, we are now reevaluating the original assumptions made based on the synthetic data. Primarily, the scanning system does not capture the curves in the road, but rather always creates a rectangular scan of the top surface. Removing the support of an arbitrary 3D point cloud allows us to greatly improve the model throughput in the order of 10-100x without losing model accuracy.
Johan Gudmundsson
CTO, Co-Founder
Learn More
More Blog Posts