Tencent Releases New Embodied VLM Model for Robotics
Tencent's Robotics X Lab, Futian Lab, and Tencent Hunyuan have released the Hy-Embodied-VLM-1.0 base model, enhancing robots' ability to understand and act within the physical environment.
Chinese technology giant Tencent has announced the release of its second-generation embodied vision-language model (VLM) base, Hy-Embodied-VLM-1.0. Developed in collaboration with Tencent Robotics X Lab, Futian Lab, and Tencent Hunyuan, the model aims to significantly improve robot autonomy and environmental interaction.
The Hy-Embodied-VLM-1.0 is designed to process and understand the physical world from a robot's perspective. It can identify objects, their attributes, depth, and spatial relationships. The model also analyzes the robot's viewpoint, functional components, and environmental affordances to plan and execute tasks more effectively.
Evaluated across 37 benchmark tasks, the model achieved an average score of 65.6. This performance is notable as its A3B scale model matches the overall capabilities of the previous generation's A32B flagship model. Key areas of strength include understanding physical states, reasoning about actions and their consequences, and temporal and adaptive reasoning.
Furthermore, the model supports long-term task planning, visual navigation, and memory functions. It can diagnose failures, perform counterfactual analysis, and dynamically replan actions based on environmental feedback. Tencent has made the model available as open-source, with code accessible via GitHub and Hugging Face.