All sessions of Transformation 2021 are now available on demand. Now look.


Last weekend, OpenAI confirmed that it had partially shut down its robotics department due to difficulties in collecting the data needed to break down technical barriers. After years of research into machines that can learn to perform tasks such as solving Rubik’s Cube, company co-ordinator Wojciech Zaremba said it is appropriate for OpenAI to shift its focus to other domains, where training data is more readily available.

In addition to the commercial motivation to eliminate robotics in favor of media synthesis and natural language processing, OpenAAI’s decision reflects the growing philosophical discussion of AI and robotics research. Some experts believe that the training system in the simulation will be enough to create robots that can complete complex tasks such as assembling electronics. Others emphasize the importance of collecting real-world data, which can provide a strong baseline.

The long-standing challenge in simulation with real data is that each scene must respond to the robot’s movements – even if they are not recorded by the original sensor. Any angle or point of view captured by a photo or video has to be rendered or simulated using predictive models dales, which is why the simulation histor is based on historically computer generated graphics and physics based rendering that somewhat crudely represent the world. .

But Julian Togelius, an AI and sports researcher and associate professor at New York University, notes that robots pose challenges that are not limited to simulation. The tires are exhausted, the tires behave differently when hot and the sensors need to be recalibrated regularly. Moreover, robots break down and slow down – and cost a pretty penny. The Shadow Dexterus Hand, a machine that OpenAIA uses in its Rubik’s Cube experiments, has a starting price in the thousands. And OpenAI had to improve arm strength by reducing its tendon tension.

“Robotics is an admirable endeavor, and I have a lot of respect for those who try to control mechanical animals,” Togelius wrote in a tweet. “But there’s no reasonable way to teach that reinforcement education, or any other episode-hungry type. In my humble opinion, the future belongs to equality. “

Training robots in simulation

Gideon Kovado, founder of Serenity, is an independent research group that works with AI to improve decision-making. Makes arguments No matter how much data is available in the real world, there is more data in the simulation – after all, data that is easier to control. The simulator can synthesize different environments and scenarios to test algorithms in specific situations. Moreover, they can randomize variables to create different training sets with different objects and environmental properties.

Indeed, Ted Xiao, a scientist at Google’s Department of Robotics, says that OpenPAI’s move to work with physical machines does not signal the end of lab research in this direction. Using techniques including reinforcement education in functions such as language and code comprehension, OpenAI can develop more capable systems that can be re-applied to robotics. For example, many robotics labs use humans with controllers to generate data to train robots. But a common AI system that understands controllers (i.e., video games) and feeds video from a camera to robotics can quickly learn to telepure.

Recent studies suggest how a simulation-first approach to robotics may work. In 2020, Nvidia and Stanford developed a technology that decomposes vision and control functions into machine learning models that can be trained separately. Microsoft has created an AI drone navigation system that can trigger appropriate actions taken with camera images. Deepmind’s scientist trained a cube-stacking system to learn from observation in a simulated environment. And a team from Google introduced a framework detail that captures animal motion capture clips and uses reinforcement education to train control policy, using adaptation techniques to disrupt dynamics in simulation, for example, mass and friction variety.

In a blog post in 2017, OpenAI researchers wrote that they believe that general-purpose robots can be created through full simulation training, followed by a small amount of self-calibration in the real world. This seems to be increasingly the case.

For AI coverage, send news tips to Kyle Wiggers – and don’t forget to subscribe to the AI ​​weekly newsletter and bookmark our AI channel, The Machine.

Thanks for reading,

Kyle Wiggers

AI staff writer

Venturebet

VentureBet’s mission is to become Digital Town Square for technical decision makers to gain knowledge about transformative technology and transactions. Our site provides essential information on data technology and strategies to guide you as you lead your organizations. We invite you to become a member of our community for:

  • Up-to-date information on topics of interest to you
  • Our newsletters
  • Gated idea-leader content and discounted access to our precious events, e.g. Transformation 2021: Learn more
  • Networking features and more

Become a member