Moving Obstacle Avoidance

Motion planning consists of finding a valid, collision-free path for a robot from a start configuration to a goal configuration. Planning in an dynamic environment is complicated by the need for constant adjustments of plans to account for moving obstacles. Yet, it is critical in real-world applications such as flight coordination, assistive robots and autonomous vehicles. In these dynamic environments, the precise future position of obstacles is often unobtainable due to either robot sensor noise or stochasticity of obstacles dynamics (such as pedestrians). It is therefore important and an active area of research to produce collision-free trajectories in the presence of these uncertainties in real-time. We focus on developing real-time motion planners with high navigation success rates for environments with a large number (up to 900) of stochastically moving obstacles and robot sensor uncertainty. We developed methods that works even when moving obstacles have strongly-interacting stochastic dynamics or completely unknown dynamics.

Strongly-Interacting obstacles with known stochastic dynamics, imperfect robot sensors

We have utilized Monte Carlo simulations to predict the position of stochastically moving obstacles in order to better inform tree-based motion planning methods with the method called Stochastic Ensemble Simulation (SES) based planning. This simulation can be done online [IROS 16, video below] for predictions of obstacles with strongly-interacting stochastic dynamics, or offline [IROS 15, video], for non-interacting obstacles. Both methods have higher success rates (up to 40% higher than comparison methods) in environments with 50 strongly-interacting obstacles and imperfect robot sensors.

Unknown obstacle dynamics

When obstacle dynamics are unknown, the navigation problem can be even more difficult. We have tackled this problem through the use of reinforcement learning where the learned goal is to arrive at the goal while maximizing distance from obstacles. This can be seen as a preference balancing task where manual derivation of optimal robot motions for these opposing preferences is difficult. PrEference Appraisal Reinforcement Learning (PEARL) [ICRA 16, video below] automatically learns near optimal motions, which solves task with opposing preferences for acceleration controlled robots. PEARL projects the high-dimensional continuous robot state space to a low dimensional preference feature space resulting in efficient and adaptable planning. PEARL can be used for dynamic obstacle avoidance robotic tasks, where an agent must navigate to the goal without collision with moving obstacles. The agent is trained with 4 static obstacles, while the trained agent avoids up to 900 moving obstacles with complex hybrid stochastic obstacle dynamics. Our results indicate PEARL has comparable success rates with state of the art methods that can require manual tuning.