Encord releases EMG-1 dataset for physical AI training
Encord has launched a 300-hour open-source dataset, EMG-1, which synchronizes muscle signals with video to improve robotic hand-movement precision.
San Francisco-based data infrastructure company Encord released a new dataset, EMG-1, aimed at advancing robotic development in physical AI. The collection comprises 300 hours of synchronized electromyography (EMG) muscle signals and first-person video recordings.
The dataset is designed to provide robotic systems with data beyond simple visual input, allowing models to interpret the physical intent and tactile mechanics behind human hand movements. By combining video with sensor-based muscle data, developers aim to bridge the gap between visual AI perception and physical execution.
This release highlights a broader industry shift toward multisensory data integration to enhance robotic dexterity. As physical AI systems transition from controlled environments to more complex tasks, the ability for machines to understand motor intent becomes a critical engineering hurdle.
The EMG-1 dataset is now available for open-source use, providing researchers and robotics engineers with a standardized set of data to train movement-based algorithms.