Single learning process for robot movements

A team of researchers from the Federal Institute of Technology in Zurich has developed a new method that allows robots to handle various movements—like climbing ladders, crawling, or getting up after a fall—using a single learning process instead of separate strategies for each task.

AI determines contact points for movement

The AI system, called Attention-based Contact Planning, uses a neural network with an attention mechanism to decide where the robot should place its body parts. It considers the terrain, the robot’s position, joint speed, and the movement goal to find the most suitable contact points.

Robot uses nine support points

The robot can use up to nine support points, including hands, elbows, knees, feet, and hips. For instance, when climbing a ladder, the AI decides where to place the hands and feet, and a separate controller then adjusts the joints and body to perform the movement while maintaining balance.

LiDAR system detects complex surfaces

To function in real environments, the system uses a LiDAR mapping system developed by the researchers. This system processes sensor data to create a terrain map and evaluates the reliability of that data. It can detect complex surfaces like narrow ladder rungs.

NVIDIA module handles data processing

The LiDAR data is processed by an NVIDIA Jetson Orin NX module, which can analyze about 20,000 LiDAR points in 30 milliseconds. The system updates the terrain map 10 times per second, allowing the robot to choose support points based on its immediate surroundings.

High success rate in simulations

In simulations, the ACP method achieved an average success rate of 92.5% on eight complex obstacle courses. This is significantly higher than the 58.8% success rate of MGMT, a standard method based on movement generation and tracking.

Simulation results need real-world testing

While the results are promising, the researchers note that these figures come from simulations and do not reflect real-world performance. Further testing in actual environments is needed to validate the system’s effectiveness.