Overview - Robotics

The goal of the robotics use case benchmark is to drive the development of robotic systems that learn in a scalable, adaptive, and privacy-preserving manner. To achieve this, we provide a structured platform for developing and evaluating methods that enable robots to learn effectively in decentralized settings.

 

 

Our benchmark focuses on federated learning approaches for acquiring manipulation skills from large-scale, decentralized, and diverse datasets. Participating teams will develop methods to train robotic agents on four different manipulation tasks using offline data. Our dataset incorporates variations in lighting conditions, textures, object appearances, and camera viewpoints, encompassing over 160,000 expert demonstrations, with a total of 16M samples. The competition is setup as a single track:

• Track 1 - Federated Imitation Learning for Robotic Manipulation: The methods will be trained within a federated imitation learning framework. In this track, the participants' objective is to learn policies in four different manipulation tasks that achieves a high level of performance.