Flying in Cluttered Environments

Part of my research in the Autonomous Systems Group at the University of Texas at Austin included building advanced obstacle avoidance capabilities for the Texas Roboticsmotion capture space so the group could test their advanced decision-making algorithms for multi-drone applications.

For this purpose, we tested various out-of-the-box algorithms, including traditional path planning algorithms from PX4. However, we did not find those to work well and, thus, developed custom deep RL algorithms on top of SOTA path planners such as egoplanner. Below are brief demonstrations of some of the developments. The simulation environment in Figure 1 is based on PEDRA.

Figure 2 demonstrates a trained RL algorithm from the PEDRA framework in the motion capture space. Note that this is an early iteration of the algorithm, and the quadcopter still exhibits trajectory planning artifacts that cause an initial jerk movement. These artifacts were fixed in later iterations.