New HY-Embodied-0.5 foundation models bridge the gap between vision-language AI and real-world robotic agents with specialized spatial-temporal perception.

HY-Embodied-0.5 comprises two variants: a 2B parameter efficient model for edge deployment and a 32B parameter powerful model for complex reasoning tasks

Models specifically designed to enhance spatial and temporal visual perception alongside embodied reasoning for prediction, interaction, and planning

Employs Mixture-of-Transformers (MoT) architecture with modality-specific computing and latent tokens to improve fine-grained visual perception for embodied agent tasks