6 October/November 2023 | Uncrewed Systems Technology Mission-critical info for uncrewed systems professionals Platform one Researchers have developed an AI system for UAVs that can beat the best human operators in competitive racing (writes Nick Flaherty). Reaching the level of professional pilots with an autonomous UAV is challenging, because the robot needs to fly at its physical limits while estimating its speed and location in the circuit exclusively from onboard sensors. A team at the University of Zurich developed the AI system for its quadrotor called the Swift that can race physical vehicles at the level of human world champions (see interview with the team’s leader on page 20). The AI system, detailed in a recent peer-reviewed paper, combines deep reinforcement learning (RL) in simulation with data collected in the physical world. The Swift competed against three human champions, including the world champions of two international leagues, in real-world head-to-head races. It won several races against each of them and posted the fastest recorded race time. The Swift consists of two key modules. A perception system translates visual and inertial information into a representation that is then processed in the second module to produce control commands. The control policy is represented by a feedforward neural network and is trained in simulation using model-free on-policy deep RL. To bridge discrepancies in the sensing and dynamics between simulations and the physical world, the team used noise models estimated from data collected on the physical system. These empirical noise models proved to be instrumental for successful transfer of the control policy from simulation to reality. A combination of learning-based and traditional algorithms is used to map onboard sensory readings to control commands. Optimising a policy purely in simulation yields poor performance of physical hardware if the discrepancies between simulations and reality are not addressed. These discrepancies are caused by the difference between simulated and real dynamics, and the noise in the UAV’s estimation of its state compared to the real sensory data. To mitigate the discrepancies, the team collected a small amount of data in the real world and used it to increase the realism of the simulator. The team recorded onboard sensory observations from the UAV together with highly accurate pose estimates from a motion-capture system while the UAV was racing through the track. During this data collection phase, the robot is controlled by a model trained in simulations that operates on the pose estimates provided by the motion-capture system. The recorded data allows the characteristic failure modes of the perception system and unmodelled dynamics that come from the environment, platform, track and sensors to be identified. It is then fed back into the augmented simulation. The Swift was tested on a physical track designed by a professional UAV racing pilot. The track has seven square gates arranged in a 30 x 30 x 8 m volume, forming a lap 75m in length. The Swift raced this track against three human champions, all using quadcopters with the same weight, shape and propulsion. The human pilots were given a week of practice on the race track. After that, each one competed against the Swift in several head-to-head races. Airborne vehicles AI wins out in UAV racing The Swift can race physical vehicles at the level of human world champions
RkJQdWJsaXNoZXIy MjI2Mzk4