Introducing Encord Data Labs

The Physical AI ecosystem is innovating faster than ever. Recent advances in robotics, world models and large-scale training are producing systems capable of more complex tasks in increasingly diverse environments. We can expect more general-purpose physical intelligence systems in the months and years ahead, and robotics is already providing early evidence of that potential. It’s an incredibly exciting time to be building AI.

The emergence of scale changes how we, as a community, think about developing Physical AI systems – particularly the amount and quality of data required. Physical environments are difficult to predict: even sophisticated models can only approximate how the world might behave, as small errors in measurement or prediction compound over time. We need systems of continual, accurate measurement in order to reflect reality. This is now the deciding factor in whether a model is truly reliable.

At Encord, we work with hundreds of AI teams developing systems for the physical world, and have seen these challenges across many different applications and modalities. The requirements for training data have evolved in tandem with model sophistication; establishing whether data accurately represents the interactions a model learns from is difficult but nonetheless critical. Our own research has grown out of these practical challenges frequently encountered by teams building and deploying Physical AI.

Today, we’re introducing Encord Data Labs, a dedicated research division focused on advancing how physical-world data is captured, understood, and evaluated. The lab brings together our work in applied research and automation, advanced data capture, and partnerships with industry leaders. We’re starting with hand tracking and new sensing modalities, where improving the accuracy of physical-world data has immediate implications for the future of robot learning.

0:00 / 0:00

0:00 / 0:00

01.Applied Research and Automation

Introducing Encord-REACH: a new hand-tracking model built for the real world.

Existing hand-tracking models often struggle with unpredictable real-world conditions, leaving gaps and errors in the annotations used to train Physical AI systems. Encord-REACH is tailored for them, improving tracking of industrial gloves, obscured fingers, fast motion and multiple hands entering the frame. We have achieved state-of-the-art 3D hand-pose accuracy in real-world environments, cutting wrist tracking error by 80% and joint tracking error by 65%. It also eliminates mistaken bystander hand swaps seen in competing models (27% on average) and reduces tracking error on industrial gloves by 30–70%. The pipeline supports additional depth, IMU and visual odometry inputs, and is now in production with leading robotics companies.

02.Advanced Data Capture

Introducing EMG-1, in collaboration with Wearable Devices

Encord and Wearable Devices are collaborating to explore how muscle signals can complement egocentric video. EMG data captures hand pose, contact, and grip force that cameras alone cannot reliably observe. Adding EMG to Encord’s market-leading egocentric data capture operation for Physical AI enables richer, more accurate human demonstration data, and more complex training for manipulation tasks.

We’re introducing EMG-1, an open dataset combining 300+ hours of synchronized wrist EMG and egocentric recordings, alongside a separate eight-channel dataset with optical motion capture. It is over 30 times larger than the next-largest public dataset of its kind. Our experiments show that combining muscle signals with video reduces median fingertip tracking error by 25%, helping recover hand movement when cameras lose visibility. We’ve also demonstrated the potential to detect finger contact directly from muscle signals. We’re excited for the research community to use EMG-1 and explore the possibilities of this new modality.

EEG, in collaboration with Zander LabsComing soon

Encord and Zander Labs are investigating how wearable EEG can capture human cognitive responses during demonstrations and robot supervision. By pairing brain signals with video and robot activity, signals such as surprise and cognitive workload can help identify valuable training moments, guide human feedback, and evaluate robot performance beyond task success alone.

03.Partnerships

NVIDIA and ANYbotics: using synthetic data to boost robotic inspection

Encord worked with NVIDIA and ANYbotics to generate synthetic examples of rare industrial anomalies, including liquid spills, 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 2.2×. Generated at a cost of just $55, ANYbotics estimates up to $78,500 in annual savings per industrial site, representing a potential return of over 1,400×.

Encord and Archetype are also evaluating how accurately AI models interpret physical signals, from heart monitors to industrial equipment, using expert-labeled data to help improve Archetype AI’s Newton physical world model. Full findings are coming soon.

The Data Labs research is fully integrated with Encord’s data development platform, syncing human data with video so that it’s ready to curate, improve and train on. Customers can bring their own teams to generate and enrich training data or work in tandem with Encord’s experts.

This is just the beginning of our ambitions for Encord Data Labs. As Physical AI continues to scale, so too will the demands on the data that underpins it. Data Labs is pushing the boundaries of what can be captured, measured, and understood in the real world – helping build the foundations of a more capable and reliable physical intelligence.