TMTPOST — AutoLife Robotics, in collaboration with Professor Cai Panpan’s research team at the Shanghai Innovation Institute, released the MINT-4B multimodal vision-language-action (VLA) foundation model on Thursday. The robotics model ranked among the top three positions in global benchmarking evaluations conducted by industry leaders including Nvidia Corp. technical experts. Technical indicators surpassed established baseline models, including OpenVLA and GR00T, due to an architecture that replicates high-level task intent rather than mimicking exact spatial trajectories. The framework utilizes a proprietary multi-scale frequency domain tokenization technology to separate top-layer operational intent from bottom-layer execution details to improve environmental adaptability. The developers have integrated the system into its AutoLife S2 humanoid robot to support commercial showroom and academic research operations. Deployments featuring integrated development and training packages have already launched across multiple regional markets within the Chinese mainland to lower operational deployment costs.