Machine Learning at Toyota Research Institute in California

I’ve spent 6 months in California working on Imitation Learning for path planning, which was later published at IROS 2020. In this work we tackle the challenge of false-positive observations from typical modern perception systems. For this we propose a novel birds-eye-view representation of the scene, that encodes the probability of an obstacle being real. This information is then used within the machine learning model. Follow this link for the published paper, or you can see the complete presentation I made for the conference.

In this video from the simulation environment CARLA, we challenge the agent by injecting wrong obstacles (empty red boxes) in front of it. The agent is able to avoid braking for wrong obstacles, but instances later gets to a full stop to avoid a collision with a real obstacle – another car.

This is a 1 minute teaser that summarizes the work!