Abstract:
Deep learning has led to significant progress in speech recognition, computer vision, natural language processing, and protein structure prediction by leveraging large data sources. Robotics, however, suffers from the lack such pre-existing repositories of behaviors to learn from; rather the robot has to learn from its own trial and error in its own specific body. This limits the diversity of scenarios that the robot can experience ahead of time, and during deployment, will inevitably encounter scenarios never seen before. To enable robots to handle these scenarios, we propose “Rapid Motor Adaptation”, a novel technique that enables robots to handle new scenarios by adapting to them in fractions of a second. Using this, we can train robots in simulation and then transfer the skills directly to robots in the real world. I will show the application of this approach to quadruped legged locomotion, biped locomotion, in-hand rotation with an anthropomorphic hand and flying quadcopters. I will then scale this idea of rapid adaptation to quadrupeds with visual sensing to achieve behaviors capable of crossing stepping stones and other challenging terrains which were beyond the reach of the blind robot.  I will also show an example of life-long learning in legged robots by continuous adaptation of perception during deployment.

Bio:
Ashish Kumar is a graduate student at UC Berkeley advised by Prof. Jitendra Malik. He works in robotics and has previously worked in long-range navigation and efficient machine learning in a few Kilobytes of RAM. Before coming to Berkeley, he was a Research Fellow at Microsoft Research India. He completed his undergraduate study at Indian Institute of Technology Jodhpur, with a major in Computer Science and Engineering. He has published in top Machine Learning and Robotics Conferences, which have received awards including a best systems paper award at CoRL. His work has also been featured in popular press including Forbes, Wall Street Journal, Washington Post, CNET, Tech Crunch, and National Geographic along with several other international venues. Webpage: https://ashish-kmr.github.io/