This research extends Chance-Constrained RRT into a receding-horizon, multi-agent planner. Each agent propagates state covariance through candidate RRT nodes, rejects branches whose estimated collision probability violates the chance bound, publishes trajectory predictions to lower-priority agents, and replans after measurement updates reduce localization uncertainty.
Classical sampling-based planners search for feasible paths in configuration space, but mobile robots also carry stochastic state estimates, process noise, prediction error, and moving-obstacle uncertainty. The CC-RRT formulation lifts collision checking into a probabilistic feasibility test: candidate nodes are retained only when the estimated collision probability satisfies the configured chance constraint.
Higher-priority agents plan first and expose time-indexed trajectory predictions. Each lower-priority planner treats those predictions as dynamic constraints alongside static geometry and moving obstacles. During execution, a lazy horizon check evaluates the next action against updated predictions; when the risk bound is violated, the planner repairs the trajectory from the current belief state.
The replay shows the lower-priority blue robot yielding to the priority green robot, taking a longer route to clear the conflict region. This behavior is the intended success case: the priority ordering, chance constraints, and receding-horizon replans produce feasible trajectories without forcing both agents through the same space at the same time.
The C++17 codebase includes a planner core, Monte Carlo collision checker, Kalman filter, runtime JSON configuration, unit tests, and CSV/JSON replay export. The public animation on this page is generated from those exported files rather than hand-drawn paths.
View the GitHub repositoryThe research was published as "On Receding Horizon Chance Constraint Motion Planning for Uncertain Multi-Agent Systems," with Vaibhav Srivastava, at ASME Dynamic Systems and Control Conference.
Read the paper
