Full-body joint control of humanoids and local execution of VLA model implemented by Google DeepMind in Gemini Robotics 2

Источник: The Robot Report

Full-body spatial control of joints of anthropomorphic platforms and adaptation of the neural network to new types of robotic bodies in just a few hours was implemented by the Google DeepMind research lab as part of the official release of the multimodal multi-agent model Gemini Robotics 2. The updated VLA class architecture combines natural language, visual perception, and direct issuance of low-level control commands to power drives without intermediate code translation through third-party translators. The key technological achievement was the ability to autonomously execute heavy neural network firmware locally directly on the robot's onboard computing modules without mandatory connection to external cloud servers.