Encord Releases Largest EMG Dataset for Robotics AI Development

SAN FRANCISCO, Oct. 20, 2026 /CourierPR/ -- Encord, a data infrastructure company for Physical AI, has released EMG-1, the largest open-source dataset linking muscle signals to first-person video. This dataset, developed in collaboration with Wearable Devices, contains over 300 hours of synchronized recordings during real-world tasks, paired with optical motion capture. It is more than 30 times larger than the next-largest public dataset of its kind.

EMG-1 addresses a significant challenge in robotics: cameras often lose track of hands when they are occluded, such as when handling objects. The dataset uses wristbands that read the forearm muscle signals behind every finger movement, known as electromyography (EMG). This technique captures what the hand is doing even when it is hidden from view. Paired with first-person video, the two streams provide a complete record of a movement that is inaccessible through a camera alone, in addition to signals like force and contact that a camera cannot capture.

Experiments showed that 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 is excited to see what the community can do with EMG-1 beyond their initial testing. Researchers can download the dataset and learn more about other initiatives at encord.com/data-labs.

Encord is also introducing Encord Data Labs, a research division working at the frontier of Physical AI data development. The lab is conducting research into how brain signals (EEG) can be used to label an operator's mental state, such as satisfaction or error response, and steer robot policies. The division's research is fully integrated with Encord's Physical AI data infrastructure, syncing human data with video in Encord so that it is ready to curate, improve, and train on.

In conjunction with the release of EMG-1, Encord Data Labs is also introducing Encord-REACH, a new state-of-the-art hand-tracking model built for real-world conditions where existing models break. Tailored for industrial conditions, the model improves tracking with industrial gloves, obscured fingers, fast motion, and multiple hands entering the frame. In Encord's evaluations, it reduced tracking errors in wrists by 80%, joints by 65%, and lowered industrial glove errors by up to 70%. It also eliminated mistaken bystander-hand swaps observed in competing models, 27% on average.

Additionally, the division is collaborating with NVIDIA and ANYbotics to generate synthetic data for rare industrial anomalies, such as liquid spills, using NVIDIA Cosmos 3 models to improve anomaly detection in robotic inspection. Using Encord's platform to curate real-world source data and validate synthetic outputs, the collaboration produced training examples for scenarios that are difficult to capture in real-world footage. Synthetic data improved recall on real-world anomalies by up to 24% and increased the examples available for model calibration by 2.2x.

In the coming months, Encord Data Labs will release more findings, including automated action segmentation for first-person video, which labels task steps from a handful of examples and significantly reduces time spent doing manual labeling. The division will also release a benchmark of frontier vision-language models showing where AI-generated video captions cannot yet be trusted for Physical AI. Expert-labeled sensor data for Archetype AI's physical world models will also be made available.