The US Researchers are trying out reinforcement algorithms techniques to automate robots. The step comes as an imperative measure to ensure the highest level of safety in navigation.However, many RL algorithm techniques focus primarily on the design of the algorithm without considering its real-time implications. This makes the performance of the UAV when embedded with RL algorithm techniques disappointing.
UAVs have limited onboard computing capabilities. Thus, if embedded with RL, algorithms take a longer time to make a prediction in real-world scenarios. As a result, it makes the UAVs unresponsive and slower, impacting the mission outcomes.
Researchersat Google Research and Harvard University have developed Air Learning. It is an open environment that will help Researchers to train the reinforcement algorithms for UAV navigation. The open-source environment is developed to enable autonomous UAVs to improve their performance results in real-world scenarios.
The Researchers‘ primary objective is to develop the open-source environment to provide foundational blocks that will help study the autonomy algorithms holistically. The air learning will allow the UAVs to train on exposure and get trained in challenging real-world settings. Researchers will train the UAVs using certain training techniques, including proximal policy optimization (PPO) algorithms and deep Q networks (DQN). The Air learning platform will allow the Researchers to assess the performance of the algorithm they have developed for the UAVs under different quality-of-flight (QoF) metrics.