Human-sized robot taught itself to walk and balance, then strolled Berkeley’s streets

UC Berkeley’s landmark Sather Gate has welcomed walkers of all kinds for more than a century, from angry protesters to students hurrying to class. Last month, though, an entirely new pedestrian passed between the gate’s granite portals.

There’s no need to run if you spot this 5-foot-tall robot ambling toward you — it’s not a militarized escapee from 2004’s “I, Robot.” Rather, this blind, teal-blue biped is the project of Berkeley scientists testing artificial intelligence research on the physical world. 

Researchers on the project aimed to develop better controls for humanoid robots by teaching it to walk, almost from scratch. Their robot spent 2023 strutting around Berkeley, taking its short, crisp steps in Edwards Stadium’s grass, in front of campus landmarks like the Campanile and alongside city roads (at present, it can’t go up or down stairs). 

In a report about the project’s findings published Dec. 14, the research team wrote that adaptability is key to making a robot useful. Current technology, though, is inflexible and limits a robot’s ability to navigate obstacles.

“Humanoid robots that can autonomously operate in diverse environments have the potential to help address labour shortages in factories, assist elderly at homes, and colonize new planets. … Here, we present a fully learning-based approach for real-world humanoid locomotion,” the report says.

Ilija Radosavovic, a Ph.D. student at Berkeley’s department of electrical engineering and computer sciences, started working on the project in a class two years ago with Bike Zhang, a mechanical engineering Ph.D. student. Radosavovic said their effort fits into research around artificial general intelligence, or the scientific pursuit of a technology that can accomplish anything a human can do — even solve new problems and reason with unfamiliar tasks.

“If you want to build a general purpose brain, you need a general purpose body, which then suggests humanoid as a platform,” he told SFGATE on Friday. The research group trained a robot built by Oregon-based Agility Robotics, designed with four joints in each arm and eight in each leg for warehouse work. 

Radosavovic, Zhang and their colleagues began by borrowing the technique of “reinforcement learning” from AI language research. They ran tens of billions of tests in a digital simulator, and an algorithm rewarded the actions that corresponded with human walking and punished the ones that didn’t. 

The training, Radosavovic said, is vaguely inspired by how a baby learns to walk. Simulated successes teach the robot what will work in the physical world. 

“Over many, many trials, in fact billions of trials, ran in simulation, the robot eventually figured out how to go from random movements and falling to balancing and walking,” he said. Guided by faculty advisers, he worked with Zhang and Tete Xiao, a recent Ph.D. graduate, on the project.

The robot doesn’t sense its surroundings like a lidar-equipped autonomous car, and it can’t look ahead. Rather, it understands how to make each step based on the precise arrangement of its limbs, where it’s just stepped and what it learned during the team’s simulations. 

Those lessons allowed the robot to better navigate terrain it hasn’t experienced, Radosavovic said. The team even threw exercise balls at the robot and shoved it with a large stick, as tests to see whether the AI learning would keep it balanced in place (it did).

“We’ve seen a lot of excitement with AI and language models and chatbots like ChatGPT, and that’s really been wonderful,” Radosavovic said. “But I feel like once AI gets to the physical world and really starts being embedded in our society, that’s going to be potentially even more exciting.”

As they trained and tested the robot, the researchers found what they called “emergent” traits they hadn’t coded in the algorithms but that are remarkably human. The robot swings its left arm forward in conjunction with its right knee (potentially because of the energy benefits, the team’s paper says) and takes smaller steps when walking down shallow slopes.

Like other robotics research, the project required huge amounts of work figuring out how to program the robot, fixing bugs and running the massive simulations. But seeing it walk, without falling (which could be a major setback if a crucial part was damaged) was “really fun,” Zhang said — as was seeing people in Berkeley crowd around the robot to take photos and videos.

“The journey is not always great,” Zhang said. “But finally when we deploy the robot in the real world, we feel very happy.”

The robot’s gait is slightly inhuman. It takes very short steps, and its arm movements are occasionally jerky. When its flat feet hit concrete and asphalt, the noise sounds more like a military march than a measured stroll, and because the robot can’t look ahead, it has to bump into an obstacle to know it can’t walk further.

Radosavovic and Zhang said that vision and more specialized arm actions are next on their agenda to study. The duo told SFGATE they aren’t yet designing with specific jobs in mind, though work automation is a major incentive for robotics development. 

They haven’t tested it as a factory worker or a delivery drone; it’s just a “general purpose” robot, so far, built to trundle around a campus. But in a batch of tests fit for a team of Berkeley students, they had the robot carry a backpack, a trash bag and a New Yorker tote.

Hear of anything happening at a Bay Area tech company? Contact tech reporter Stephen Council securely at [email protected] or on Signal at 628-204-5452.