Sign In

Communications of the ACM

ACM TechNews

Virtual Obstacle Courses Help Real Robots Learn to Walk

View as: Print Mobile App Share:

The 4,000 ANYmal simulated robots were trained using reinforcement learning.

Credit: Nvidia

Researchers at Switzerland's ETH Zurich and chipmaker Nvidia developed an army of more than 4,000 simulated doglike robots and used them to train an algorithm to control a real-world robot's legs.

These "ANYmals" can navigate slopes, steps, steep drops, and other challenges in a virtual environment.

The algorithm was transferred to a real-world, four-legged robot, which could navigate stairs and blocks but encountered difficulties at higher speeds due to what the researchers said were inaccuracies in how the robot's sensors perceive the real world versus the simulation.

The researchers said running the simulations on Nvidia's chips instead of general-purpose chips allowed them to train the robots in less than a hundredth of the time it typically takes.

From Wired
View Full Article


Abstracts Copyright © 2021 SmithBucklin, Washington, DC, USA


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account