Successfully taught humanoid robot 'Figure 02' to walk like a human, and a creepy video of a large number of robots marching
American robotics company
Natural Humanoid Walk Using Reinforcement Learning
https://www.figure.ai/news/reinforcement-learning-walking
Watch eerie video of humanoid robot 'army' marching naturally, thanks to a major AI upgrade | Live Science
https://www.livescience.com/technology/robotics/watch-eerie-video-of-army-of-humanoid-robots-marching-naturally-thanks-to-a-major-ai-upgrade
Figure believes that humanoid robots capable of human-like movements could be a solution to the growing labor shortage in warehouses and other workplaces. However, while humanoid robots can perform advanced movements such as backflips and breakdancing , they are not good at 'walking naturally like a human being' and can only walk in an awkward manner that makes them seem robotic.
The paradox that it is harder to give a computer the perception and motor skills of a one-year-old than it is to teach it how to win at chess, known as Moravec's paradox , is that computers excel at problems that require complex calculations and large data sets, but lack the real-world experience based on millions of years of evolution.
So Figure decided to teach its humanoid robots to walk naturally using a technique called reinforcement learning, in which an intelligent agent learns through trial and error and optimizes its behavior based on reward signals. By running reinforcement learning through a high-fidelity physics simulator, Figure explains, it can simulate years' worth of data in just a few hours.
You can see Figure 02, which has actually learned how to walk through reinforcement learning, walking in the video below.
Introducing Natural Walking - YouTube
Figure 02 walking through the entrance of an office building.
Although it does feel a little awkward compared to a real human, it's not as awkward as you might imagine when you hear 'how a robot walks.'
Let's actually compare Figure 02 (V1), which was trained using the previous method, with Figure 02 (V2), which was trained using the new method.
Comparing the gait of V1, which is displayed lightly, with the gait of V2, which is displayed normally, we can see that there are clear differences in the degree of bending of the knees and the way the feet move.
V2, which came later, quickly overtook V1.
To teach a humanoid robot how to walk, Figure uses reinforcement learning in a GPU-powered physics simulator.
Thousands of Figure 02s were simulated in parallel.
Each Figure 02 had its own unique physical parameters and was exposed to a variety of scenarios it could encounter.
The model built by reinforcement learning can be applied to Figure 02 without any additional adjustments, and the video shows 10 Figure 02s marching without any fine-tuning or changes.
'While these initial results are exciting, we believe they only hint at the full potential of our technology. We are committed to extending our learned policies to handle the full range of human-like scenarios that robots may face in the real world,' Figure said in a blog post.
Related Posts:
