FRONTIER

Anotation

Computer vision is beginning to see a paradigm shift with large-scale foundational models that demonstrate impressive results on a wide range of recognition tasks. While achieving impressive results, these models learn only static 2D image representations based on observed correlations between still images and natural language. However, our world is three-dimensional, full of dynamic events  and causal interactions. We argue that the next scientific challenge is to invent foundational models for embodied perception – that is
perception for systems that have a physical body, operate in a dynamic 3D world and interact with the surrounding environment.

FRONTIER addresses this challenge by means of:

1. developing a new class of foundational model architectures grounded in the geometrical and physical structure of the world that seamlessly combine large-scale neural networks with learnable differentiable physical simulation components to achieve generalization across tasks, situations and environments;
2. designing new learning algorithms that incorporate the physical and geometric structure as constraints on the learning process to achieve new levels of data efficiency with the aim of bringing intelligent systems closer to humans who can often learn from only a few available examples;
3. developing new federated learning methods that will allow sharing and accumulating learning experiences across different embodied systems thereby achieving new levels of scale, accuracy, and robustness not achievable by learning in any individual system alone.

Breakthrough progress on these problems would have profound implications on our everyday lives as well as science and commerce with safer cars that learn from each other, intelligent production lines that collaboratively adapt to new workflows or a new generation of smart assistive robots that automatically learn new skills from the Internet and each other enabled by the advances from this project.