Unlike existing robots on the market, such as Boston Dynamics’ pointwhich uses an internal map to move around, the robot uses cameras alone to guide its movements in the wild, said Ashish Kumar, a graduate student at UC Berkeley and co-author of the paper describing the work; Robot Learning Conference next month. Other attempts to use camera cues to guide robot movement have been limited to flat terrain, but they managed to get their robot to walk up stairs, climb rocks and jump over gaps.
The quadruped robot is first trained in a simulator to move around different environments, so it has a rough idea of walking in a park or going up and down stairs. When it’s deployed in the real world, vision from a single camera in front of the robot guides its movements. The robot learned to adjust its gait to navigate things like stairs and uneven floors using reinforcement learning, an artificial intelligence technique that allows systems to improve through trial and error.
Deepak Pathak, an assistant professor at Carnegie Mellon University, said removing the need for an internal map makes the robot more robust because it is no longer constrained by potential errors in the map.
Jie Tan, a research scientist at Google, who was not involved in the study, said robots have a hard time translating a camera’s raw pixels into the kind of precise and balanced movements needed to navigate their surroundings. This work, he says, is the first time he’s seen such amazing maneuverability in a small, low-cost robot.
The team has achieved “a breakthrough in robot learning and autonomy,” said Guanya Shi, a researcher at the University of Washington who studies machine learning and robot control, who was also not involved in the study.
Akshara Rai, a research scientist at Facebook AI Research who works in machine learning and robotics but was not involved in the work, agrees.
“This work is a promising step toward building such sentient legged robots and deploying them in the wild,” Rai said.
However, while the team’s work helps improve how the robot walks, it won’t help the robot figure out where to go in advance, Rai said. “Navigation is important for deploying robots in the real world,” she said.
More work is needed before the robot dog can hop around the park or fetch things around the house. While the robot can learn depth through its front-facing camera, it can’t handle things like slippery ground or tall grass; it can get bogged down in puddles or get bogged down in mud, Tan said.