Situation Principles • Objects in lanes we plan to cross are more interesting. • Closer objects are more interesting. • Moving objects are more interesting. • Objects moving towards us are more interesting.
State ‘Parsing’ • Each tracked object i will have a probability of being in each lane l, pil. • They will also have a distribution over speed and distance to next waypoint, Pil_distance(x) and Pil_speed(v). • These will be half guessed based on the output of the Dynamic State Estimator and the RNDF.
Planner Interaction • The ‘planner’ will need to provide input to the interpreter. • One way is some kind of query/reply. • The Interpreter will find the relevant parts of the state and estimate the probability of a conflict with the intention. • It will then provide a useful response to the Planner.
Query/Reply: Road-follow • Query: What is the road-follow situation at segment 5.2, position 15 m., timestamp t? • Reply: safe, lane center/right/left edges in dead reckoning frame, minimum speed and distance of moving car ahead, grid map of relevant region showing drivable road surface. • For unknown situations, it might be possible to just put a phantom object at the unknown location until it is known thus causing the planner to slow down and prepare to stop.
Query/Reply Lane Change • Query: what is lane change situation at lane 3.4, position 23.4 m, timestamp t? • Reply: Safe, new lane center/edges in dead-reckoning frame, speed and position of car ahead and behind in new lane, grid map of drivable road expected to be safe. • Lack of knowledge might cause an unsafe reply. This might then change to safe as time goes on.
Query/Reply: Left Turn • Query: What about a left turn from exit waypoint x to entry waypoint y at time t? • Reply: Safe, the exit and the entry in the dead-reckoning frame, along with a grid map of the intersection in that frame showing the cells that are expected to be safe during the turn. • Unsafe might be due to lack of information. The reply will give the position of the last safe part of our lane, (for nosing out to look).
Implementation • We could parameterize each common situation, pre-calculate the decision surfaces and put in a table. (Fast but not so general). • We need to detect when the world is not as expected, cars behaving erratically. This should trigger some kind of safe behavior? • Need to break out of deadlocks. • Deciding that we do not know enough is hard.