Seminar | March 21 | 10-11 a.m. | Hybrid. In person: BWW8019; zoom: link below
Tete Xiao
Electrical Engineering and Computer Sciences (EECS)
In the deep learning era, the remarkable progress of computer vision has been driven by learning representations on task-specific data and annotations. However, this paradigm faces mounting challenges in the real world, especially in robotics, due to constantly changing environments and safety concerns. These challenges render the current approach unscalable, limiting its ability to adapt to new scenarios and tasks. In this talk, I will discuss the key ingredients for learning scalable representations for computer vision and robotics, which include models, data, and learning objectives. I will start with a project that evaluates the optimizability of large vision neural networks through a case study of vision transformer (ViT) models. Next, I will present an inductive bias-free learning objective for contrastive self-supervised representation learning. Reducing inductive biases is crucial for learning general representations that can be utilized in various downstream tasks. Following this, I will present works that learn and deploy visual representations through self-supervised visual pre-training from diverse, in-the-wild visual data for real-world robotic tasks, showcasing the significance of scalable visual representations for real-world robot learning. Lastly, I will present a learning-based approach for real-world humanoid locomotion, demonstrating the potential of learning robotic representations from simulation.
Trevor Darrell, trevor@eecs.berkeley.edu, 415-690-0822