dhruvms/pushplan

Repository Architecture

• The Planner class runs the outer loop for WHCA*.
• Planner initialises a CollisionChecker object shared between itself and all agents - it populates the immovable obstacles right from the get go.
• The idea is that CollisionChecker will also be updated with the current positions and trajectories of movable agents for dynamic obstacle collision checking.
• Each Agent will run it's own A* search in (x, y, t) to compute a path.
  • For the object-of-interest, we'll implement a windowed (limited expansion) search.
  • For the other movable objects, we should do the same, or we could experiment with running a BFS to the same depth, and returning the "most open" solution?
  • Initially, the abstract space distance will just be the naive heuristic (Euclidean or Manhattan distance). Eventually we'll implement RRA*.

Helper Functions

• geometry.hpp contains a bunch of functions to aid the collision checker.
  • Functions to compute circle-rectangle and rectangle-rectangle intersections.

Domain Notes

• All agents except the OOI cannot exit table bounds.
• For the time being, agents cannot rotate in place.
  • Next step: allow in-place rotation
  • After that: loop in the simulator and change actions to force vectors.
• If you look at a scene image from the simplan/data/clutter_scenes folder, the positive X-axis goes right, positive Y-axis goes up.
• Planner (WHCA*) terminates when the OOI has a collision-free path from its initial start to goal, and all other agents have collision-free paths (where to does not matter).
• For any agent, even if start == goal, we still run the search till depth w, i.e. the is_goal check should return true if a state at depth w+1 is being expanded.