DeepCybo, the first embodied intelligence company incubated by Beijing Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence, is building foundation models for general physical intelligence. Their approach centers on training with egocentric human demonstration data, capturing how humans naturally interact with objects to give robots true understanding of physical manipulation.
The challenge? Collecting high-quality hand motion data that's accurate enough to train policies for dexterous manipulation. Camera-based tracking systems struggle with occlusion when hands grip objects, fingers overlap, or objects block camera view. For embodied AI training pipelines, data quality directly determines model performance.
MANUS data gloves provide DeepCybo with drift-free, occlusion-free hand motion capture at 25 degrees of freedom, tracking every joint angle across the thumb and fingers in real-time. The electromagnetic field (EMF) based system maintains millimeter-level precision throughout extended data collection sessions without requiring line-of-sight or recalibration. This means cleaner and higher quality training data, essential for learning generalizable manipulation policies.
“After comparing IMU, optical, and vision-based solutions for our robotic foundation model development, MANUS gloves delivered the most reliable and robust performance.”
Xuguo He, R&D head at DeepCybo
Beyond data collection, DeepCybo uses MANUS gloves for real-time teleoperation of dexterous robotic hands, validating their trained policies through direct human control. The sub-40ms latency enables responsive control where operators feel direct connection to the robot hand,making it possible to execute complex bimanual tasks and evaluate policy performance against human baselines.
