This little robot can go almost anywhere.

Researchers from the School of Computer Science at Carnegie Mellon University and the University of California, Berkeley have designed a robotic system that allows an inexpensive and relatively small robot to climb and descend stairs almost at its height; traversing rocky, slippery, rugged, steep and varied terrain; cross gaps; climbing rocks and curbs; and even operate in the dark.

“Empower little robots climbing stairs and managing a variety of environments is crucial for developing robots that will be useful in people’s homes as well as for search and rescue operations,” said Deepak Pathak, assistant professor at the Institute of Robotics. “This system creates a robust and adaptable robot that could perform many everyday tasks.”

The team put the robot through its paces, testing it on uneven stairs and slopes in public parks, challenging it to walk on steps and slippery surfaces, and asking it to climb stairs that, for its height, would resemble a human leap. a barrier. The robot adapts quickly and masters difficult terrain by relying on its vision and a small on-board computer.

The researchers trained the robot with 4,000 clones of it in a simulator, where they practiced walking and climbing over difficult terrain. The speed of the simulator allowed the robot to acquire six years of experience in a single day. The simulator also stored the motor skills he learned during training in a neural network that the researchers copied from the real robot. This approach required no manual engineering of the robot’s movements, a departure from traditional methods.

Most robotic systems use cameras to create a map of the surrounding environment and use this map to plan movements before executing them. The process is slow and can often fail due to inherent fuzziness, inaccuracies or misperceptions in the mapping stage that affect subsequent planning and movement. Mapping and planning are useful in systems focused on high-level control, but are not always suited to the dynamic demands of low-level skills like walking or running over difficult terrain.

The new system bypasses the mapping and planning phases and routes vision inputs directly to robot control. What the robot sees determines how it moves. Even the researchers don’t specify how the legs should move. This technique allows the robot to react quickly to oncoming terrain and traverse it efficiently.

Because there’s no mapping or planning involved and the movements are trained using machine learning, the robot itself can be inexpensive. The robot used by the team was at least 25 times cheaper than available alternatives. The team’s algorithm has the potential to make low-cost robots much more widely available.

“This system uses vision and feedback from the body directly as input to send commands to the robot’s motors,” said Ananye Agarwal, an SCS Ph.D. machine learning student. “This technique allows the system to be very robust in the real world. If it slips down the stairs, it can recover. It can go to unfamiliar environments and adapt.”

This direct aspect of vision to control is biologically inspired. Humans and animals use vision to get around. Try running or balancing with your eyes closed. Previous research by the team had shown that blind robots — robots without cameras — can conquer difficult terrain, but adding vision and relying on that vision dramatically improves the system.

The team also looked to nature for other elements of the system. In order for a small robot – less than a foot tall, in this case – to be able to climb stairs or obstacles almost to its height, it has learned to adopt the movement that humans use to cross tall obstacles.

When a human needs to lift their leg high to climb a ledge or obstacle, they use their hips to move their leg sideways, called abduction and adduction, which gives them more clearance. The robot Pathak’s team-designed system does the same, using hip abduction to tackle obstacles that trip some of the more forward legs. robotic systems on the market.

The movement of hind legs by four-legged animals also inspired the team. When a cat moves through obstacles, its hind paws avoid the same objects as its front paws without the benefit of a pair of nearby eyes. “Four-legged animals have a memory that allows their hind legs to follow the front legs. Our system works the same way,” Pathak said. The system’s onboard memory allows the rear legs to remember what the front camera saw and maneuver to avoid obstacles.

“Since there is no map, no planning, our system remembers the terrain and how it moved the front leg and translates it to the back leg, doing it quickly and perfectly,” said Ashish Kumar, Ph.D. student at Berkeley.

The research could be a big step towards solving existing challenges facing legged robots and introducing them into people’s homes. The paper “Legged Locomotion in Challenging Terrains Using Egocentric Vision”, authored by Pathak, Berkeley Professor Jitendra Malik, Agarwal and Kumar, will be presented at the upcoming Robot Learning Conference in Auckland, New Zealand.