Zero-shot, One-shot, and Few-shot Robot Learning

  • Date and time: Wednesday 3 April 2024, 2pm
  • Location: In-person and online
    ISA-135 Seminar Room, Institute for Safe Autonomy (Map)
  • Admission: Free admission, booking not required

Event details

Robot learning offers a flexible and general approach for robots to learn new skills. However, the typical methods used by the field, such as Reinforcement Learning and Behavioural Cloning, are highly inefficient. In this talk, I will present research from the past year in our lab which deviates from these typical trends, and aims towards more practical robot learning without huge data requirements. For example, we have studied and developed (1) methods which can work zero-shot — using pre-trained Vision-Language Models without requiring any task-specific data, (2) methods which can work one-shot — requiring only a single demonstration for each new task, and (3) methods which can work few-shot — requiring 3 or 4 demonstrations per task and subsequently enabling generalisation across object shapes. As a common theme through all of these projects, we are studying general methods which work in real-world environments with real camera observations, and which can teach robots typical everyday tasks, such as scooping up an egg from a pan, pouring from a teapot into a cup, and opening a lock with a key.

 Join us on Zoom if you can't attend in person

About the speaker

Dr Edward Johns

Dr Edward Johns is the Director of the Robot Learning Lab at Imperial College London, and a Senior Lecturer (Associate Professor). His work lies at the intersection of robotics, computer vision, and machine learning, and he and his team are currently studying visually-guided robot manipulation. He received a BA and MEng in Electrical and Information Engineering from Cambridge University, and a PhD in visual place recognition from Imperial College. Following his PhD, he was a post-doc at UCL, before returning to Imperial College as a founding member of the Dyson Robotics Lab with Prof Andrew Davison, where he led the robot manipulation research. In 2017, he was awarded a prestigious Royal Academy of Engineering Research Fellowship for his project "Empowering Next-Generation Robots with Dexterous Manipulation: Deep Learning via Simulation", and then in 2018 he was appointed as a Lecturer and founded the Robot Learning Lab. He established Imperial College's first course on Robot Learning, which he currently teaches at graduate level. In 2022 he received Imperial College's President's Award for Outstanding Early Career Researcher. Externally, he is on the advisory board for a number of robotics startups, including Karakuri and Muddy Machines, and from 2021 to 2022 he spent a year as Head of Robot Learning at Dyson in a part-time role.