Visual Learning for Collision Avoidance in a Simulated Environment

We demonstrate a simulated robot in a three-dimensional, texture-mapped graphical environment that learns from optical ow calculations to avoid collisions with walls. The robot has no preprogrammed notion of divergence or left-right asymmetry, but by associating the values of visual variables leading up to the moment of each collision with the fact and particulars of that collision, the robot gradually approximates these functions and improves its performance. The credit assignment problem associated with reinforcement learning is greatly reduced in this continuous domain, permitting a simple nearest-neighbor learning algorithm. It is hoped that this technique may be extended to other mobile robot problems for which appropriate functions are not as easily discerned from rst principles. Further , the use of artiicial visual environments is advanced as a tool for judging the feasibility of vision-based learning algorithms before committing them to real robot platforms.