In January 2019, I joined the simulation team at drive.ai for a 4-months internship,
where I worked as a software engineer.
Specifically, I focused on the problem of using static annotations and HD maps
to improve the creation and behavior of simulated dynamic agents.
Leveraging static annotations and HD maps, we created a semantically rich lane-level
graph, that simulated agents can use to navigate the map.
This lane graph is comprised of three layers: topological, metrical, and semantic.
Topological and metrical layers were extracted from HD maps and static annotations,
while the semantic layer was constructed from annotations only.
Our lane graph, superimposed to an HD map, allowed us to:
- speed up the generation of new synthetic scenarios by using our graph to guide
simulated agents from an arbitrary point A to an arbitrary point B in our map;
- reduce the amount of labor required to define these trajectories and their parameters;
- smooth the produced trajectories by representing them using piecewise polynomials;
- create simulated agents that can detect and react to common events along their
trajectories, such as, changes in speed limit, traffic lights, stop signs, etc.;