China trains robots to jump like cats to explore asteroids

Inspired by cats' ability to turn and land, a research team at the Harbin Institute of Technology (China) used reinforcement learning (RL) – a type of artificial intelligence (AI) – to train robots to adjust their posture in the air when jumping over rough, low-gravity surfaces on asteroids.

A Chinese research team trained a four-legged robot to adjust its posture and land like a cat to move on the surface of an asteroid. (Photo: SCMP)

Unlike traditional systems that rely on specialized but heavy stabilization hardware, the robot uses a “model-free” control system to move its four legs in a coordinated motion. This allows the robot to adjust its tilt and reorient its direction of travel in mid-air, the researchers report in the Journal of Astronautics.

The research addresses a key challenge with robot jumping when moving on asteroids, where the environment has low gravity and even a slight imbalance in leg forces can cause the robot to spin uncontrollably, land unsuccessfully, or bounce off the surface entirely.

“In the low gravity environment of asteroids, robots experience long periods of free fall during each jump. It is important to use this time to adjust the deflection caused by the jump, to ensure a safe landing or to change the rotation angle to adjust the direction of movement.”, the research team said in the report.

“A microgravity simulation platform was designed and built, thereby verifying the effectiveness of this jumping method through experiments on a quadruped robot prototype,” the research team added.

Asteroids are remnants of the formation of the solar system and hold the key to deciphering its origins. They are also rich in resources such as platinum and other rare metals, which could aid future space exploration and industrial applications.

Challenges on the asteroid surface

So far, space agencies in Europe, Japan and the US have successfully landed spacecraft on asteroids to retrieve samples, but none have deployed rovers capable of long-term surface exploration.

Traditional wheeled rovers, like those used on the Moon and Mars, face challenges in asteroid environments because the weak gravity, typically just a few thousandths of Earth's, does not provide enough traction for wheels to operate effectively.

To address these limitations, scientists have proposed using jumping robots for future missions, but that presents a new set of challenges.

Each time it jumps, the robot stays in the air for about 10 seconds or so, long enough for the unbalanced leg forces to cause the robot to spin uncontrollably or even bounce off the surface and drift into space.

The Harbin team used RL to train the robot in a virtual simulation. Over seven hours, the AI ​​learned from its experimental mistakes and refined its movements to land stably. The robot’s AI system demonstrated the ability to adjust its orientation, including pitch (leaning forward or backward), tilt (leaning sideways), and yaw (rotation angle), in just a few seconds.

For example, when launching forward at a large tilt of up to 140 degrees, the robot can stabilize its posture within 8 seconds. It can also rotate in mid-air up to 90 degrees to change direction of movement.

Robots are trained using reinforcement learning. (Photo: SCMP)

To validate the effectiveness of the system, the researchers built a microgravity simulation platform that allows the robot to “float” on a nearly frictionless surface.

Although limited to two-dimensional motion, the experiments confirmed the effectiveness of the system and reinforced the results from simulations, the team said.

Additionally, the scientists found that the process requires very little computing power from the robot. The system's lightweight and energy-efficient design makes it particularly suitable for deep space exploration missions.

In the future, this system could have a wide range of applications, from scientific exploration to resource mining on asteroids. However, the team said more research is needed to improve the AI's ability to adapt to diverse terrains and environments.

Hoa Vu(Source: SCMP)