Driving through narrow passages, dense lane changes and unsignalled crossings calls for proactive but interpretable planning, which end-to-end learned methods struggle to deliver, according to a paper on arXiv.
Avinash Singh and colleagues present SIGMO, an interaction-aware motion planner built on polynomial sigmoid gates. The gates weigh how much each surrounding agent matters, so the planner scales to many agents and adapts as their behavior changes.
The authors say the approach stays in harmony with surrounding traffic, avoiding both freezing and excessive caution.
