25届未来精英-Robotaxi Core Planner

Use uncertainty signal from perception/prediction for planning ego behavior with safety and assertiveness in autonomous driving

Planner is responsible for generating a feasible, comfortable, and safe trajectory given perception and prediction output. Beyond that, we need ego vehicle to be legible and thus determined when driving on road.

One challenge is upstream is not always perfect, for example, perception can not have 100%/100% PR, nor can it provides an uncertainty singal that is completely interpretable. This applies to prediction as well, as prediction is basically just a likelihood distribution. So planner needs to plan in existence of upstream uncertainty, to ensure ego behavior is both smooth, safe and efficient. 

The other challenge is self driving vehicle(SDC) drives in a highly interactive environment, ego's behavior impacts other agents’, and other agents' behavior impacts ours too. How to elicit meaningful interactions between ego and other agents in the real world is the second challenge. 

The other challenge is human-like behavior in SDC. There are traffic rules that only applies to SDC, but does not applies to social vehicles. E.g., social vehicles can drive slightly above speed limit, but by no means can SDC do so. This poses another challenge for SDC that it has to drive like human in a social environment while not being able to follow the same exact rules as other social agents.