Genesis AI's Novel Approach to Robot Training

Vivian Sun has been around the block in autonomous driving. There were stints at trucking companies TuSimple and Waabi, where she respectively served as VP of business and chief commercial officer, followed by a stint as Amazon’s head of autonomous driving.  

In March, Sun pivoted to a new — though related — field, joining Genesis AI. The move was a gamble of sorts. While still not fully mature, we’re more than a decade into the world of big self-driving bets. In spite of a seemingly endless glut of demo videos, physical AI is a nascent technology by comparison.  

It can be difficult to ignore some glaring parallels with the early days of self-driving. If past is, indeed, prologue, the landscape is set to shift dramatically.  It’s precisely these unknowns, Sun contends, that make the early-stage startup world so attractive. 

“I really enjoyed the beginning phase, the zero to one,” says Sun. “That's why I picked the last 10 years of autonomous driving, not the next 10 years. I think this is exactly the moment we have for robotics. We're seeing true generalization capabilities. The robots can like us. There can be a pre-trained brain that can do so many things more things than ever before. And we don't need to program for anything. You give it data, you learn, and there you go.” 

Sun adds that she doesn’t expect physical AI generalization to take as long to mature as autonomous driving before it.  “Obviously, we're not there yet, but we’re seeing the light at the end of the tunnel,” she says. “That's such an important spark in many people. This is the right moment for robotics, and I think the robotics industry is not going to take 10 years [to mature].” 

For one thing, Sun says, there’s more low-hanging fruit in terms of useful, real-world robot deployment that can go toward training and kickstarting the data flywheel. For another, many of these tasks don’t require the same level of reliability or safety needed in order to put an autonomous system on public roads.  

“Autonomous driving is just such a safety critical system,” says Sun. “For example, the trucks we built at Wabi and TuSimple are heavy 18-wheelers running on the highway at 60 mph. There's just no way you can stop in the middle of it. The safety implication is so high. In order to deploy, you need to get to 99.99%. That's probably not even enough safety, reliability, and we go through such thorough analysis in terms of automotive standard, all that. Whereas with robots, we're going through the generalization phase, if you do one, two, or three tasks really, really well, there's already huge market for you.” 

Sun is quick to add that — as with other physical AI startups — Genesis is aiming for generalization, rather than a few goals. Her point, however, is that real-world training begins far earlier than such generalization is achieved.  

One trick Genesis AI has up its sleeve is a glove designed to collect data as humans do work. The form factor allows the workers to do their jobs as they would normally, presenting what the company calls a “zero domain gap” that can be transferred to an end effector designed to accurately mimic the human hand.  

Whether the five-finger hand is an ideal robot gripper remains a hot topic of debate in the robotics world, but for Genesis’ purposes, at least, the design makes a lot of sense. After all, it’s the form factor humans are using to log the data in the first place. 

“The majority of the reason for the justification for the dexterous hand is for the data transfer,” says Sun. “That's the only way to unlock the unlimited human data. Otherwise, you're going to create some form factor, some end factor that's like mid-tier or not future-proofed. We can also argue the world is built for a human hand, so having the ultimate form factor will also help. You just have much more dexterity and capabilities with five finger hands.”