Multi-Robot Path Planning in Dynamic Environments Using RRT* Algorithms

MRPP=Multi-Robot Path Planning in dynamic envs requires collision-free, time-opt traj gen under moving obsts & inter-agent interactions. RRT*=Rapidly-exploring Random Tree Star, an asymptotically opt incremental sampling-based MOT=Motion Planning Algo. Base RRT builds random tree by sampling C-space (config space), extending toward rand pts via steer fn, but yields subopt solns. RRT* improves via rewiring: after node insertion, checks K-near nodes for lower-cost parents, updates edges if cost-reduction found → drives convergence to opt traj over time. In MRPP, direct RRT* appl faces curse of dim: C-space dims scale as O(n×d), n=robots, d=per-robot DOF. Soln: decoupled appr — plan seq or in parallel w/ conflict res. PAR-RRT*: prioritized planning, robots plan one-by-one w/ higher-priority paths reserved as dyn obsts. Limit: priority bias. ORCA+RRT*: hybrid local avoid w/ global RRT* roadmap. For dyn envs, ST-RRT*=Space-Time RRT*, extends C-space with time dim → C_free(t) → enables time-param traj gen avoiding moving obsts. Time discretization critical; too coarse → miss coll, too fine → comp load. DRRT*=Dynamic RRT*, updates tree via local regen when obsts move; uses change detection in env to trigger partial replan. For MRPP, deconfliction via CBS=CBS (Conflict-Based Search) layered on RRT*: high-level RRT* gen init paths, low-level CBS res conflicts (vertex, edge, temp) via split constraints. Scalability: n>10 → comp burden. Soln: Dec-RRT*=Decentralized RRT*, each robot runs local RRT* w/ comms for pose/intent sharing. Comm loss → risk coll. Use VO=Velocity Obstacles or HRVO=HRVO for local correction. RRT*-CD: Collision Detection enhanced via BB=Bounding Boxes or TSDF=Truncated Signed Distance Fields for fast C-space eval. In practice, kinodynamic constraints (vel/acc/jerk) enforced via state-space ext: RRT* in X=R^6 (pos, vel, acc) → ensures dyn feasibility. Diff: high-freq replan in dyn envs → latency issues. Use warm-start: init tree from prev soln. Adaptive sampling: bias samples near obsts or conflict zones via GMRS=Guided Multi-Robot Sampling. MR-RRT* with MPC: embed RRT* in receding horizon; replan every Δt using pred obst traj (from KF/EKF). Pred error → risk. Use robust variants: RR-RRT*=Risk-RRT*, model obst motion as Gaussians, minimize coll prob. Eval metrics: SOC=Sum of Costs, MF=Makespan, Coll Rate, Comp Time. Benchmarks: SIM-UR, MAPF-STD, or custom ROS/Gazebo sim. Real-robot tests: Khephera, TurtleBots, or UAV swarms. Pitfalls: 1) Overhead from comms in Dec-RRT*, 2) Deadlocks in PAR-RRT* w/ cyclic dep, 3) High false-pos in CD due to sensor noise, 4) Subopt due to early term in RRT*, 5) Scalability limits in dense envs. SOTA: 1) RRT*-SMART: uses machine learning (MLP/CNN) to guide sampling toward feasible regions, 2) G-RRTP: game-theoretic RRT* for adversarial dyn envs, 3) Neuro-RRT*: DL-based latent space biasing, 4) Q-RRT*: quantum-inspired sampling for faster cov. Tools: OMPL, ROS MoveIt!, STORM, MARPLE. Best practices: hybrid global/local (RRT* + DWA), use TSDF for env rep, implement anytime RRT* for real-time, validate with Monte Carlo sim. Future: integration with LLM for intent pred, neuromorphic RRT* for low-latency, sim2real via domain randomization.