San Francisco AI company Encord released EMG-1, an open dataset that pairs more than 300 hours of wrist muscle signals with first-person video, including people playing Jenga, to help robots learn to handle objects like humans. It was unveiled Oct. 20 at Encord's Long Horizon Physical AI Summit at Fort Mason.
Why it matters: Cameras lose track of hands when an object blocks the view. Muscle signals keep recording what each finger is doing, which can make robot training data more accurate.
How it works
Wristbands read the electrical signals from forearm muscles, a technique called electromyography (EMG), while a head-mounted camera records what the person sees. Encord built the dataset with Wearable Devices.
"Jenga is a very difficult problem for a robot because you have to rely on multiple different sensor inputs," Encord co-founder and co-CEO Ulrik Hansen told USA TODAY.
The numbers
- 307 hours of field recordings plus a 10.4-hour lab collection with optical motion capture, according to Encord.
- Combining muscle signals with video cut fingertip tracking error by 25% compared with video alone, and by 35% during short gaps in visual tracking, the company said.
- Encord says it is more than 30 times larger than the next-largest public dataset of its kind.
Encord said the data will be released on Hugging Face and hopes it will help train humanoid robots, autonomous vehicles and drones.