Shield AI Inc. — Department of Defense SBIR Phase II: AF193-CSO1

Description

AI for unmanned robotic systems entails the development of state-of-the-art algorithmic approaches for autonomy and intelligence. Using simulation, Shield AI can increase the velocity of AI algorithm development. Shield AI’s high-fidelity simulation framework models the real world and simulates sensor data, which enables the accelerated and parallelized training of robotic intelligence, leading to generational capability leaps measured in weeks, not years. The broader robotics community from academia and industry has been using reinforcement learning (RL) for fast optimization of executive planning and behaviors with positive results. RL is an approach to machine learning that defines a reward function and learns a policy that maximizes the reward. The benefit of framing a problem as an RL problem with an appropriate reward function is that the policy can be learned off of many examples rather than by manually tuning the parameters of the policy.  This effort seeks to apply RL to the development of autonomous behaviors for a team of air-launched effects (ALEs) performing the military mission of environment monitoring within our high-fidelity simulation. The capability development process with simulation and RL is extremely fast, which allows for thousands of iteration and makes for a very robust transition to real-world platforms.