Overview - Robotics
The objective of the privacy preserving federated reinforcement learning robotics competition is to develop privacy-preserving solutions for robotic platforms acting on household environments. In particular, we seek federated reinforcement learning approaches for the training of robotic agents in simulation environments.
The participating teams will create methods to train robotic agents on a mobile pick-and-place task using reinforcement learning on provided household environments. Due to the sensitive nature of the data, the training procedure should uphold privacy guarantees, using a federated-learning set-up. The competition is structured in two different tracks:
Track 1 - Federated Reinforcement Learning: The methods will be trained within a federated reinforcement learning framework. In this track, the participants' objective is to learn policies in a mobile pick-and-place task that achieves a high level of performance, while avoiding collisions with humans and other objects in the environment to ensure safety.
Track 2 - Federated Reinforcement Learning + Differential privacy: In this track, in addition to training over distributed data, we add another layer of protection for the clients. In particular, we employ a red-blue team approach, where blue team participants develop learning methods that protect the identity of the clients, and red team participants attempt to identify which client environments were employed in the training of the blue-team solutions.
Competition/Challenge TBA for Q1 2024