GPU-accelerated real-time motion planning for safe human-robot collaboration

Abstract

Robotic automation has a significant role in industry over decades. As the demand for complex tasks increases, there has been a recent anticipation for robotic automation in human-robot collaborative environments, leading to the introduction of commercial collaborative robots. However, current robot controllers reduce the speed of robots to secure the safety of humans, which results in conservative behavior and lower performance with collaborative robots. A challenge in this thesis is to maximize the productivity of collaborative robots while ensuring safety, aiming for productivity of robots under collaborative situations comparable to that of traditional robots.

In our first work, we introduce a rapid trajectory smoother, primarily to enhance productivity. Existing real-time path planners lack the smoothing post-processing step -- which is crucial in sampling-based motion planning -- resulting in the trajectories being jerky, and therefore inefficient and less human-friendly. Our rapid trajectory smoother, based on a shortcutting technique, leverages fast clearance inference by a novel neural network and can consistently smooth a trajectory for a 6 DoF robot within 200 ms on a commercial GPU. A comparison shows that our smoother is faster than the state-of-the-art method and the smoothed trajectory is more efficient than the original jerky trajectory even when considering the time required for smoothing.

Subsequently, we propose a time-optimal safe path tracking algorithm, with a particular focus on ensuring safety. Our path tracking algorithm is formulated based on Time-Optimal Path Problem based on Reachability Analysis (TOPP-RA) and proven to provide the fastest control policy for controlling a robot to track a given path. Our method guarantees the safety of human operators in the sense that the robot will collide only when the robot has a zero velocity, in accordance with ISO safety standards. We also demonstrate the application of our method in a 6-DoF industrial robot scenario.

Another challenge is that, to achieve true time-optimality in safe path tracking, it is crucial to have precise distances between obstacles and a robot at waypoints along an executing path. However, existing methods for computing distances between a robot and obstacles are either too slow for real-time applications, or inaccurate for achieving time-optimality. Thus, we propose a batched distance checker for time-optimal safe path tracking. Our method can evaluate distances of a trajectory in less than 1 millisecond on GPU at runtime, making it suitable for time-critical robotic control. We experimentally demonstrate that our method can navigate a 6-DoF robot earlier than a geometric-primitives-based distance checker in a dynamic, collaborative environment.

Throughout this thesis, we emphasize the performance of our algorithms and their implementations. Since our focus is on industrial applications, algorithm performance is critical for the practicality of our methods. Parallelization plays an important role in achieving high performance, especially with the widespread and powerful GPUs. Therefore, in addition to explaining the proposed algorithms, we develop and benchmark our GPU-accelerated implementations.

We hope that this thesis will pave the way for further development and application of human-robot collaboration both in industry and beyond.