Using Unity's Machine Learning Agents on Arm

Unity’s Machine Learning Agents (ML-Agents) is an open-source intelligent game character kit that requires no experience. The framework can be used to train single- or multi-agent scenarios with reinforcement learning, imitation learning and other deep learning algorithms for use in games and apps, on mobile and PC. It sits atop Unity Barracuda, and is available on GitHub.

On Windows on Arm we used ML-Agents to train a boss character to fight against in our Dr Arm game, with reinforcement learning. The agent trained with adversarial self-play, with a gym of fights against a constantly improving version of itself allowing the agent to slowly learn better attacking techniques. With this we investigate how Arm CPUs can be best used for ML in games.

Ongoing research is looking at multi-agent scenarios, to see how the workloads scale with number of agents and with interactions between them. ML-Agents is a great use-case for CPU inference on both mobile and Windows on Arm.

There will be a large zip file with the required resources made available on this session page the week before the workshop. Attendees are also encouraged to install Unity and to create an account on GitHub. The workshop will require an Android-based mobile phone with a USB connector to the development machine.

By downloading the workshop file below, you agree to be bound by the terms and conditions stated in the EULA.