Meet Cassie, the Usain Bolt of robots

On a crisp May day in Oregon, it wasn’t an Olympian but a robot called Cassie that broke a Guinness world record for the 100-meter dash.

The robot, which researchers say resembles an “ostrich without a head,” started the day with a few stumbles, but ultimately prevailed — running 100 meters in 24.73 seconds, slower than Usain Bolt’s record-setting 9.58 seconds, but still a Guinness world record for a bipedal robot, Oregon State University announced last week.

Cassie’s roughly 40 supporters were elated, cheering when it crossed the finish line. Its success was a seminal moment in robotic history, they said. Cassie’s speed and agility, honed by artificial intelligence training, showed that bipedal robots could maneuver in taxing real-world situations while maintaining balance, a problem that has plagued designers in the past.

The race built on Cassie’s 2021 successful completion of a 5K in roughly 53 minutes, which first showed Cassie could stay upright for long periods. It was also the capstone to about five years of work by engineering and machine learning researchers at Oregon State University and a spinout company, Agility Robotics, paving the way for more advanced designs.

 “This is the first big step to humanoid robots doing real work in the real world,” said Alan Fern, an artificial intelligence professor at Oregon State University who helped train Cassie. “Because [now], we can get robots to robustly move around the world on two legs.”

For decades, scientists, entrepreneurs and engineers have clamored for two-legged robots. In the 1960s, Japanese researchers created rudimentary prototypes of bipedal machines. In the past decade, engineers at MIT and the California Institute of Technology have tried to do the same. Last week, Tesla CEO Elon Musk debuted a two-legged humanoid robot, Optimus.

 But two-legged robots have always faced problems, researchers said, namely losing balance and falling down.

To solve that, Fern teamed up with Jonathan Hunt, an Oregon State University professor and co-founder of Agility Robotics, to train bipedal robots using machine learning and neural networks, which are algorithms that mimic the way a human brain works.

< Watch the video at the source page. >

 The research is funded in part by the National Science Foundation and the Defense Advanced Research Projects Agency (DARPA), a secretive government organization behind creations such as the internet.

Since 2017, the team has been training Cassie how to walk properly, using algorithms to reward the robot when it moves appropriately. “This is all inspired by Pavlovian psychology,” Fern said. “It just learns to anticipate these rewards and do the right thing.”

Once the team got the remote-controlled robot working well in simulation, the next step was seeing how it would handle real-world environments, where surfaces are uneven, friction can change and a robot’s mass can shift.

In 2021, when the team had Cassie run a 5K, it learned a few things. The robot was being “too stompy,” Fern said, and researchers started rewarding the robot when it smoothed out its gait. With this year’s successful 100-meter dash, the team is moving on to the next step: putting a torso and head on Cassie. (Agility Robotics is working on one called Digit.)

Fern said that will bring engineers one step closer to humanlike robots that could one day move around packages in warehouses, build homes or provide care for the elderly in homes.

But such advances come with their own challenges.

Humanoid robots with heads propped onto Cassie’s leg design will need the peripheral vision to navigate tricky terrain. “Now, Cassie has to look around the world,” he said, “understanding what objects are there and not running into them.”

The robot will also have to identify an object as something to pick up, and then be intelligent enough to do it the way a human would. (For example, Fern said, if a robot was asked to put boxes in a room, it should load boxes in from back to front.)

Most important, these robots must walk with intention. “When you’re in the real world,” Fern said, “sometimes you do have to actually pay attention to where you step.”

Still, engineering experts said it will be an uphill climb to replace humans with robots.

Nancy J. Cooke, a professor of human systems engineering at Arizona State University, noted that robots are getting very good at doing things like running or kicking a soccer ball. The harder part is creating a machine that interacts with humans in a natural way.

“What they’re lacking is really complex cognition,” Cooke said. “There’s still a deep understanding of humans that’s needed to interact with humans that they don’t have.”

Cooke also said it’s laudable that robots like Cassie are advancing the robotics industry, but it seems unnecessary to build machines that simply replicate what humans do. It might be more worthwhile, she said, to create robots that can do things humans cannot.

“Why [do] we want to rebuild ourselves?” she asked. “I think it’s a sci-fi thing, but other than entertainment value — I think it’s an overkill.”

Source