American academics design tiny robots that can go anywhere

Researchers at Carnegie Mellon University's School of Computer Science and the University of California, Berkeley have designed a robotic system that enables a low-cost robot with relatively small legs to climb and descend stairs close to its height, traverse rocky, slippery, Uneven, steep and varied terrain. Bridge chasms, peel rocks and curbs, and even work in the dark.

"Giving small robots the ability to climb stairs and handle a variety of environments is critical to developing robots that are useful in people's homes, as well as in search-and-rescue operations," said Deepak Pathak, an assistant professor at the Robotics Institute. Robots that can perform many everyday tasks."

The team put the robot to the test, testing it on uneven stairs and hillsides in public parks, challenging it to step over stepping stones and slippery surfaces, and asking it to climb stairs because it was as tall as a human jumping over an obstacle. The robot relies on its vision and a small onboard computer to quickly adapt and master challenging terrain.

The researchers trained the robots with 4,000 clones in a simulator, where they practiced walking and climbing on challenging terrain. The speed of the simulator allows the robot to gain six years of experience in one day. The simulator also stored motor skills learned during training in the neural network, which the researchers replicated on the real robot. This approach does not require any manual engineering of the robot's motion -- unlike traditional methods.

Most robotic systems use cameras to create a map of their surroundings and use this map to plan movements before execution. The process is slow, and problems often arise due to ambiguity, inaccuracies, or misunderstandings inherent in the mapping phase, affecting subsequent planning and movement. Mapping and planning are useful in systems focused on high-level control, but are not always suitable for the dynamic demands of low-level skills, such as walking or running on challenging terrain.

The new system bypasses the mapping and planning stages and directly routes visual input to the robot's control. What the robot sees determines how it moves. Even the researchers didn't specify how the legs should move. This technology allows the robot to quickly respond to oncoming terrain and move through it efficiently.

Because no mapping or planning is required, and machine learning is used to train movements, the robots themselves can be low-cost. The robot the team used is at least 25 times cheaper than existing alternatives. The team's algorithm has the potential to make low-cost robots more widely available.

Ananye Agarwal, a doctoral student in machine learning at SCS, said: "The system uses vision and feedback from the body directly as input to output commands to the robot's motors. This technique makes the system very robust in the real world. If it slips on the stairs, It can recover. It can go into unknown environments and adapt."

This direct vision of control is biologically inspired. Humans and animals use vision to move. Try running or balancing with your eyes closed. The team's previous research has shown that blind robots (those without cameras) can conquer challenging terrain, but adding vision and relying on it can greatly improve the system.

The team also looked to nature for other elements of the system. For a small robot less than a foot tall to climb stairs or obstacles close to its height, it learned to adopt the motions humans use to step over tall obstacles. When a person has to lift their legs high to climb a rung or obstacle, it uses the hips to move the legs out of the way, called abduction and adduction, which gives it more room. The same is true for the robotic system designed by Pathak's team, which uses hip abduction to overcome obstacles that hold back some of the most advanced legged robotic systems on the market.

The movement of the quadruped's hind legs also inspired the team. When a cat moves through an obstacle, its hind legs avoid the same objects as its front legs, without the help of a nearby pair of eyes. "Four-legged animals have a memory that enables their hind legs to track their front legs. Our system works in a similar way," Pathak said. The system's on-board memory enables the rear legs to remember what the front camera sees and maneuver to avoid obstacles.

"Because there is no map, no planning, our system remembers the terrain and how it moves its front legs, and converts that to its rear legs, and does it so quickly and perfectly," said Ashish Kumar, a Ph.D. student at Berkeley. This research could be a big step toward solving existing challenges with legged robots and bringing them into people's homes.