Hafiq’s Deep Reinforcement Learning-based robot is doing well in crowd navigation. The DRL model was trained in one crowd setting and tested in different unseen crowd settings. In simulation, we have tested with up to 12 moving obstacles at a speed close to the robot. In real-world, we have tested with 4 obstacles at a speed half of the robot’s speed. In all test settings, the robot successfully navigated to the goal position. It wasn’t easy to teleop the many obstacle robots in real-world test as the obstacles were colliding with each other under cumbersome manual control; the robot was avoiding them! So, we had to limit the number of obstacles in real-world test.
The main idea is to give the robot with the perceived risk of the moving obstacles (crowd) by including the collision probability of the most dangerous obstacles in the observation space. We have submitted the paper to IROS 2023. More info and source code will be shared soon.
Project page: https://ailab.space/projects/personal-robots/#robot-navi-rl