Monte Carlo Tree Search with Velocity Obstacles for safe and efficient motion planning in dynamic environments

TL;DR

This work tackles online motion planning for robots in densely populated, dynamic environments with partial obstacle information. It introduces a VO-enhanced MCTS framework that prunes unsafe actions using only obstacle positions and maximum speeds, enabling safe and efficient planning with modest computational budgets. The key contribution is integrating VO into both the tree search and rollout phases to dramatically reduce the required simulations while maintaining safety, demonstrated in challenging 10×10 m scenes with up to 40 obstacles where it outperforms NMPC and DWA. The approach is training-free, scalable, and practically impactful for real-time navigation in crowds and cluttered spaces.

Abstract

Online motion planning is a challenging problem for intelligent robots moving in dense environments with dynamic obstacles, e.g., crowds. In this work, we propose a novel approach for optimal and safe online motion planning with minimal information about dynamic obstacles. Specifically, our approach requires only the current position of the obstacles and their maximum speed, but it does not need any information about their exact trajectories or dynamic model. The proposed methodology combines Monte Carlo Tree Search (MCTS), for online optimal planning via model simulations, with Velocity Obstacles (VO), for obstacle avoidance. We perform experiments in a cluttered simulated environment with walls, and up to 40 dynamic obstacles moving with random velocities and directions. With an ablation study, we show the key contribution of VO in scaling up the efficiency of MCTS, selecting the safest and most rewarding actions in the tree of simulations. Moreover, we show the superiority of our methodology with respect to state-of-the-art planners, including Non-linear Model Predictive Control (NMPC), in terms of improved collision rate, computational and task performance.

Figures (5)

• Figure 1: The velocity space for a robot moving in an environment with one obstacle. The blue region ($CC_i$) denotes the collision cone corresponding to the obstacle, a circle representing $\pazocal{B}(\bm{p}_i, r_i+r_R)$. Thus, $\bm{v}_r$ (the relative velocity of the robot to the obstacle) is infeasible, and another velocity (e.g., $v$) must be selected, such that $v \in \pazocal{V} \setminus CC_i$.
• Figure 2: a) The robot ($\bm{p}_R$) is out of the extended ball $\pazocal{B}(\bm{p}_i, r_2)$ (yellow circle); b) The robot is inside the extended ball, but still not colliding with the physical ball of the obstacle $\pazocal{B}(\bm{p}_i, r_i)$ (pink circle). Blue circle: $\pazocal{B}(\bm{p}_R, r_1)$; red cone: $CC_i$ delimited by tangent angles $[\alpha_1, \alpha_2]$ (Algorithm \ref{['alg:cc']}, Line 8).
• Figure 3: Snapshot of the actual environment with obstacle radius $r_i = 0.2m$ and robot radius $r_R = 0.3m$ (a); and radii $r_1, r_2$ as of Algorithm \ref{['alg:cc']}. Blue circle: agent. Black circles: dynamic obstacles. Red cross: goal.
• Figure 4: Comparison between MCTS_VO_TREE, MCTS, NMPC, DWA and VO_PLANNER ($m$ is the number of MCTS simulations)
• Figure 5: Results of the ablation study and comparison between MCTS_VO_TREE, MCTS, MCTS_VO2 and MCTS_VO_ROLLOUT.