PhD in Vision-based Robot Learning with World Models

BeschreibungFacebook Twitter LinkedInNavigating complex unstructured environments requires a deep understanding of the robot's surroundings. Where is it safe to navigate? What actions are too risky to take? How can the robot quickly adapt from experience collected during operation? To answer these questions, novel algorithms that combine machine learning, control, and vision are required.Autonomous navigation based on onboard sensors and computation has made tremendous progress over the last decade (e.g., NASA-JPL Mars helicopter, Tesla autopilot, Boston Dynamics Atlas, UniTree G1). However, the performance of these systems is still far from humans in of agility, versatility, and robustness: "agility" will allow the robot to increase its productivity, "versatility" to adapt to new environments and continually learn from new data, and "robustness" to succeed at any task.The goal of the position will be to research deep learning algorithms (imitation learning, reinforcement learning, differentiable simulation) with world models to train sensorimotor policies that can navigate robots better than humans, end to end, by mapping visual inputs to control commands. Experimental platforms range from flying robots to legged robots and cars. Your research topics will revolve around learning and adapting in the real world, learning from offline data (e.g., YouTube videos), and generalist, multi-task learning (learning behaviors that generalize to multiple tasks and scenarios). The specific topic will be identified together with Prof. Davide Scaramuzza. #J-18808-Ljbffr