BeingBeyond Unveils Being-H0.8, the First Implicit Tactile World-Action Model Trained on 500,000 Hours of Video

BeingBeyond has unveiled Being-H0.8, the first implicit tactile world-action model built from human video data. The model incorporates tactile modality into a latent space representation, expanding the previous Being-H0.7 visual world model. It uses a Mixture of Transformers (MoT) to predict interaction outcomes, a slow-fast action expert for execution and real-time adjustment, a Universal Tactile Encoder to unify tactile signals, and TopoHand to align human and robot action spaces. The training data pool exceeds 500,000 hours of first-person video, cleaned and processed into the UniHand-3.0 dataset. To address the lack of tactile labels in video, BeingBeyond developed TactoHand to infer contact and proximity from visual data, supplemented by real pressure glove data. The model was tested on dual-arm robots performing tasks such as retrieving objects from a bag, writing with a brush, squeezing toothpaste, and picking up chips, demonstrating its ability to handle delicate manipulations requiring tactile feedback. BeingBeyond was founded in May 2025 by Lu Zongqing, an associate professor at Peking University, and has progressed through three stages: proving feasibility of human video training (Being-H0 to H0.5), scaling up (H0.5 to H0.7), and now improving data quality (H0.8).