Jenga Teaches Robots Human Touch: Encord's Leap in AI Training with Sensors

A San Francisco startup, Encord, is making big strides in robot training by using the popular game Jenga, along with head-mounted cameras and muscle sensors. They've just released the EMG-1 open dataset, which contains over 300 hours of real-world recordings, allowing robots to learn delicate hand movements, even when cameras can't see the fingers directly. This breakthrough helps teach robots the kind of precise touch normally only humans have. The challenge for 'Physical AI' – robots that work in the real world – is that there isn't enough good training data, unlike the huge amounts of text available for chatbots. Encord is tackling this by 'manufacturing' data: they hire people to physically show robots tasks while sensors record everything from muscle signals to brain waves, in partnership with German neuroscience startup Zander Labs. This focus on 'dense annotation' aims to give robots high-quality information about how much pressure to apply and where to move, which is crucial for complex tasks like playing Jenga. This new method is set to improve how we train humanoid robots and other AI systems that need to interact physically with our world. By giving robots the ability to learn through human actions and sensory feedback, we could soon see machines that are much more adaptable and skilled, moving beyond simple, repetitive jobs. The next big thing will be watching how quickly these highly trained robots can take on unpredictable, real-world environments.