When robots need a brain, their creators turn to Silicon Valley. Most of the AI tech that drives advanced robotics originates in Bay Area labs. At last month’s ReWork Deep Learning for Robotics Summit in San Francisco, researchers from Silicon Valley AI labs and institutes discussed their latest work and how it is being used to teach robots. Synced was onsite to bring you an inside look at their work.
Google: Self-supervised learning
From recipes to repairs, humans can quickly learn how to do something by watching it done in a video. Google Brain believes it can realize this ability in robots.
Google Brain Research Engineer Corey Lynch discussed a self-supervised approach for learning representations and robotic behaviors entirely from unlabeled multi-perspective task-related video data. Google researchers studied how this representation can be used in two robotic imitation settings: imitating object interactions from videos of humans, and imitating human poses. The research was first published in the 2017 paper Time-Contrastive Networks: Self-Supervised Learning from Video.
Researchers introduce a time-contrastive approach, which uses multi-view metric learning to represent an image frame in an embedding space. The trained robots can learn how to recognize the same movement from different viewpoints, while also distinguishing different movements even if they look similar.
The representation can then be used as a reward function in reinforcement learning algorithms, training robots repetitively until they finally know how to precisely perform the target task. Demonstration tasks included for example pouring water from a pitcher.
Navigation is a key robotics research interest, as it is believed effective navigational skills will help to mature robot technologies generally. Navigation is more complicated than other challenges due to the many factors that need to be taken into consideration in dynamic real-world environments. Today’s simulated environments for robot navigation training are oversimplified.
Facebook AI Research (FAIR) Research Scientist Georgia Gkioxari presented House3D, a rich, extensible and efficient environment that contains 45,622 human-designed 3D scenes of visually realistic living environments ranging from studio apartments to multi-storied houses. The environments are equipped with a diverse set of fully labeled 3D objects, textures and scene layouts.
Facebook AI bots trained in House3D achieved a performance boost of eight percent over bots trained with only raw RGB images.
BAIR: Robot acrobat
The Berkeley Artificial Intelligence Research (BAIR) Lab had the most dramatic approach: teaching robots to spin-kick like a karate black belt or backflip like an acrobat. Jason Peng, a first year PhD student at UC Berkeley, explained how a simulated robot can become a stuntman.
Peng co-authored the paper DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills which demonstrates a Reinforcement Learning (RL) technique that enables simulated characters to replicate highly dynamic physical movements learned from data collected from video of human subjects.
Using a method known as Reference State Initialization (RSI) and video of a human acrobat, the character can learn skills from any state of moves, such as the inflection point of a flip, and RSI can allow the character to know which states will result in high rewards even before it has acquired the proficiency to reach those states. Early Termination (ET) meanwhile is a standard practice for RL researchers to stop simulations that lead to failure. Like a human athlete, the character is driven to strive for excellence, and the results are nearly indistinguishable from the human reference subjects.
The two-day ReWork Summit featured 30 speakers from industry and academia and 200 leading technologists and innovators, and ran June 28-29 at the South San Francisco Conference Center.
Journalist: Tony Peng | Editor: Michael Sarazen