Boston Dynamics and Toyota Research Institute (TRI) on Wednesday unveiled plans to bring AI-based robotic intelligence to their electric humanoid robot Atlas. This collaboration leverages work done by TRI on large-scale behavior models (LBMs) that operate along similar lines to the more general large-scale language models (LLMs) behind platforms such as ChatGPT .
Last September, TechCrunch visited TRI's Bay Area campus to get a closer look at the institute's work on robotic learning. In research revealed at last year's Disrupt conference, the institute's director, Gil Pratt, described how an overnight training taught a robot to perform household tasks like flipping pancakes with 90% accuracy. I explained how I was able to do it.
“In machine learning, until recently, there was a trade-off that worked well but required millions of training cases,” Pratt explained at the time. “When you're doing physical work, you don't have time to do that many things. The machine breaks down before you get to 10,000. Now it looks like you need dozens of them. Dozens of reasons. , because you need to have variety in your training cases. But sometimes it's less than that.”
Boston Dynamics is a good fit with TRI in terms of hardware. Spot manufacturers have played a role in software and AI to enhance their systems, but the working methods required to teach robots to perform complex tasks with full autonomy are completely new. It's a different beast.
Gill Pratt speaks at TC Session: Robotics 2017 Image credit: TechCrunch
“There has never been a more exciting time for the robotics industry, and we look forward to working with TRI to accelerate the development of general purpose humanoids,” Boston Dynamics CEO Robert Player said in a statement. There is. “This partnership is an example of two companies with strong research and development bases working together to tackle many complex challenges and build useful robots that solve real-world problems.”
Boston Dynamics unveiled its electric Atlas design in April, finally putting its large hydraulic humanoid namesake to rest. The robot hasn't been seen much since then, but in August TechCrunch was able to obtain a short video of the robot doing push-ups. As with Atlas' first video, the quick push-up demo was a good demonstration of the robot's remarkable strength.
Boston Dynamics' major competitors in the humanoid robot space, including Agility, Figure, and Tesla, have largely chosen to build their AI teams in-house. The partnership between Boston Dynamics and TRI is particularly interesting given that the organization is operated by Hyundai and Toyota, which are direct competitors in the automotive field.
Image credit: Boston Dynamics
Meanwhile, Boston Dynamics has its own research spinout, The AI Institute (formerly known as The Boston Dynamics AI Institute). Although run by Boston Dynamics founder and former CEO Mark Lybert, the institute maintains its independence from Boston Dynamics. We are also a very young organization and are still strengthening our team. TRI has reduced its investment in hardware.
The goal of all of this is a truly versatile machine. That is, a system that can learn and do everything a human can do, and perhaps more. We've seen robot hardware evolve to the point where that level of sophistication is possible, but anything approaching general intelligence is much harder to solve.
Indeed, the advent of SDKs for systems has dramatically expanded the range of tasks that robots like Boston Dynamics' Spot can perform, but true artificial general intelligence is still a long way off.