Encord releases EMG-1, largest open dataset linking muscle signals to robot AI

What EMG-1 contains

EMG-1 was developed in collaboration with Wearable Devices. It uses wristbands that read the forearm muscle signals behind every finger movement, a technique called electromyography (EMG). This lets the dataset capture what a hand is doing even when it is hidden from view.

The recordings were made during real world tasks and are synchronized with optical motion capture. Researchers can download the dataset at encord.com/data-labs.

Why cameras fall short

Encord explains that autonomous robots learn to manipulate objects by training on real-world video of people completing tasks. The problem is that cameras lose track of hands when they are occluded, most often by the object being handled.

EMG is used to produce accurate hand-pose labels on training data where the camera fails, since it continually records the electrical activity that drives each finger. Paired with first-person video, the two streams give a complete record of a movement, in addition to signals such as force and contact that a camera cannot capture.

In Encord's experiments, combining EMG with video reduced fingertip tracking error by 25% compared with video alone, and resulted in 35% lower error during short gaps in visual tracking. EMG alone also recognized pinch-like finger movements.

Encord Data Labs and further releases

EMG-1 launched together with Encord Data Labs, a new research division focused on Physical AI data development. Encord says it covers Applied Research & Automation, Advanced Data Capture, and Partnerships. EMG is described as the first of several biological signals the lab plans to bring into robot training data. It is also researching how brain signals (EEG) can label an operator's mental state, such as satisfaction or error response, and steer robot policies.

Two other releases came alongside the dataset:

• Encord-REACH: a hand-tracking model built for real-world conditions, tailored for industrial settings with gloves, obscured fingers, fast motion and multiple hands in frame. In Encord's evaluations it reduced tracking errors in wrists by 80%, joints by 65%, and industrial glove errors by up to 70%. It also eliminated mistaken bystander-hand swaps seen in competing models, 27% on average.
• Synthetic data findings with NVIDIA and ANYbotics: the companies generated synthetic examples of rare industrial anomalies, including liquid spills, with NVIDIA Cosmos 3 models. According to the release, synthetic data improved recall on real-world anomalies by up to 24% and increased the examples available for model calibration by 2.2x.

Encord plans more findings in the coming months: automated action segmentation for first-person video, a benchmark of frontier vision-language models showing where AI-generated video captions cannot yet be trusted for Physical AI, and expert-labeled sensor data for Archetype AI's physical world models.

What it means for robotics teams

Encord Co-Founder and Co-CEO Eric Landau said that "real-world model performance comes down to the data going in," and described EMG-1 and the new hand-tracking model as a solution to "one of the biggest problems facing robotics teams today." Encord says it works with hundreds of companies in the field, including Toyota, Boston Dynamics and Zipline.

The division's research is integrated with Encord's Physical AI data infrastructure, which syncs human data with video so it is ready to curate, improve and train on. Customers can bring their own teams to generate and enrich training data or work with Encord's experts. The announcement was made at Encord's inaugural Long Horizon Summit in San Francisco, where Data Labs demonstrated EMG-1 and its hand tracking model live.

For researchers, the open release offers a public resource for training on hand movements that cameras alone cannot see. The performance figures come from Encord's own experiments and evaluations.