Sophisticated sensor technology can be misled by the way lidar beams reflect off of something as mundane as a puddle.
Why it matters: Semi autonomous and fully autonomous vehicles are largely expected to reduce accidents, but aspects of the tech stack can introduce risk. Simulation is one way to discover scenarios that may fool sensors and train perception algorithms accordingly.
What's happening: Large-scale sensor simulation enables a fast and reliable path towards perception system development by testing thousands of scenarios under hundreds of different conditions, from the size and distribution of the puddles to the speed and direction of vehicles and environmental backscattering and absorption.
Between the lines: Simulations have pointed out the limitations of sensor technology, like lidar, for example.
- Road conditions influence how lidar beams are reflected after hitting the road. Depending on the reflective qualities of a surface, a lidar receiver may or may not receive the reflection of a fired laser beam.
The bottom line: A perception algorithm has to decide what to do with received points, determining whether they are real or whether they are false positives. Extensive simulation is needed to generate enough data to train perception algorithms to catch these kinds of edge cases.
Amin Aghaei is a sensor simulation engineer at Metamoto, an AV simulations company.