What we need in the Golden Age of Robotics

While bipedal humanoid robots are super-popular right now from manufacturers like Apptronik, Amazon (in partnership with Agility Robotics), Sanctuary AI, Figure.ai, Tesla and Fourier Intelligence, it’s likely that robot innovation will diversify rather than stagnate. is consolidated. the next few years as we enter an emerging golden age of robotics. At least, that’s my takeaway from a conversation with some of the top scientists and engineers at Boston Dynamics, which pioneered humanoid robots with Atlas and continues to build numerous robots in various form factors.

Investments in robotics fell last year, but the global robotics market is projected to grow almost 15% annually through 2032.

Increasingly, robots will run like dogs, fly like birds, wiggle like snakes, swim like fish, roll like cars and — yes — walk like people. There’s no right form factor, says Boston Dynamics senior technical director Alex Perkins.

“The right form factor is whatever the right form factor is for the application you’re really targeting,” he told me on a recent TechFirst podcast. “We tend to gravitate towards things we know: the human form, the animal form, and there’s a great utility and flexibility in those forms.”

But, he adds, robot designs can be extremely specific to individual tasks: even a form of writing and stationary like a regular dishwasher.

Interestingly, what we think of as the best and most adaptable forms of robotics, like the two-legged, two-armed human form that we use to great effect for so many different kinds of tasks, are not always the most the good ones.

Take moving boxes to a warehouse.

“The first box mover we ever did was with Atlas, and as we started to understand that market more, we moved away from this walking robot,” says Boston Dynamics’ Mike Murphy, senior technical director and principal systems engineer. “And we get to Stretch where we have static balance. So the evolution of the form to actually fit the application is really one of the most beautiful parts of the pinnacle of the golden age of robotics.”

Atlas is Boston Dynamics’ two-legged humanoid robot that plays parkour. When building a warehouse and logistics operation robot, Boston Dynamics first evolved the Handle form factor, a two-wheeled robot with a single ostrich-like neck/arm. Eventually, however, engineers switched to Stretch, a newer robot with a four-wheeled base that can carry much more battery power with it while working just as well in warehouses, which tend to have flat concrete floors.

“When you compare [Stretch] against the ostrich machine, it was five times faster as soon as we produced the first prototype,” says Perkins. “We’re like, oh yeah, that’s the thing for this app. And we’ve moved cases much faster since then.”

Interestingly, Stretch does not have fingers and hands like robots like Tesla’s Optimus, Agility Robotics’ Digit or Sanctuary AI’s robots do. Instead, Stretch uses suction cups that allow it to grab and move boxes that may be tightly packed together or awkwardly shaped.

For getting around the world or a workplace, feet can make sense, of course, especially on uneven ground. But two legs may not always be the solution here either. Marco da Silva, Boston Dynamics vice president and general manager for Spot, the company’s dog-like robot, says he gained a new appreciation for a four-legged form a few years ago in an uncontrolled outdoor environment.

“The idea was to show Spot and Atlas walking together in the woods, and Spot was basically just circling Atlas,” he says. “Our robot was walking down the street and doing an admirable job, but Spot was just running through the woods, climbing hills … just doing laps.”

Any hiker who has done long steep sections of trail at speed, only to be passed by a happy dog ​​going four times faster, knows exactly how that felt.

There may also be a cost advantage to the quadrupedal form factor: bipedal locomotion requires many more actuators, or robotic “muscles.” Think about it: the hips, knees, and ankles all require articulation for effective bipedal form. Murphy, the lead systems engineer, says that while a dog-like robot might require only three actuators per leg, a human-like robot might need 15.

This translates to 30 total actuators for a humanoid robotic leg versus only 12 for a dog form.

The humanoid form factor persists because we tend to think that our build has made us great general purpose machines. But this is a bit of a mistake, says da Silva.

“If you look at something like a humanoid robot, it’s easy for people and investors to imagine that it can, you know, do anything a person can do,” he says. “But then when you get into a specific application and understand the market better, you start to see that there are some trade-offs you can make to make the car more reliable, to make it more performant, to make it cheaper. “

No one has yet made a humanoid robot that is fast and has a long battery life, for example.

Another key area that needs innovation in robotics is training. Apptronik has started using LLM to train its robots: using natural language to tell a robot what to do and have the AI ​​system translate this into discrete steps and actions. This is essential for using robots at home and in work situations.

For example, asking a robot to slice a tomato requires a large amount of knowledge: where it might be, what a tomato is, what kind of knife might work on a tomato, where that knife is, what used a cutting board … and more. And that doesn’t even include how to cut it, or what to do with the slices, or how to clean up afterwards.

LLMs can help, says Murphy.

“In the warehousing space, a lot of warehouse management systems basically use voice commands for a picker to say, ‘Go to this bay and get me five boxes,'” he says. “The level of software integration that would have to come in and the number of APIs that you would have to cut to integrate with all these different warehouse management systems could be immense. Or you can just take that very simple command that one understands instantly and kidding, you’re done.”

In terms of form factor, it doesn’t look like the Boston Robotics team is hot on humanoids. The main limiting factor is how much battery power you can carry.

Ultimately, it’s still a goal, but Murphy suggests it remains a stretch that requires significant technological progress before it’s feasible.

“If the end of all form factors is humanoid robots, I don’t think anyone here will argue, but anything that develops in pursuit of a humanoid robot, I think represents great technological progress for robotics. as a whole”, he says.