While autonomous vehicles hold the promise to become more reliable and safe than their human counterpart, long tails of unfortunate events are statistically bound to happen (e.g., jaywalkers, hardware and software failures,...).
In these emergency situations, one needs to take complex decisions in a split second to effectively mitigate the collision severity (or to avoid it). In this project, we want to augment a nominal set of precomputed emergency trajectories with a few additional learned candidates.
The advantage is that the learning can happen offline (e.g., simulation), where for each emergency situation one can compute a globally optimal solution without time constraints.
Then online, one would use the learned map (raw sensor data → evasive trajectory) to enrich the precomputed set of trajectories to be evaluated in an emergency situation
While autonomous vehicles hold the promise to become more reliable and safe than their human counterpart, long tails of unfortunate events are statistically bound to happen (e.g., jaywalkers, hardware and software failures,...).
In these emergency situations, one needs to take complex decisions in a split second to effectively mitigate the collision severity (or to avoid it). In this project, we want to augment a nominal set of precomputed emergency trajectories with a few additional learned candidates.
The advantage is that the learning can happen offline (e.g., simulation), where for each emergency situation one can compute a globally optimal solution without time constraints. Then online, one would use the learned map (raw sensor data → evasive trajectory) to enrich the precomputed set of trajectories to be evaluated in an emergency situation
- Design and implement the infrastructure to generate learning data
- Design and implement the training algorithm
- Deploy and evaluate the solutions on real autonomous racing karts
- Design and implement the infrastructure to generate learning data - Design and implement the training algorithm - Deploy and evaluate the solutions on real autonomous racing karts
