Google’s DeepMind Robotics Team has unveiled a series of new technologies aimed at advancing the field of robotics. The new systems, AutoRT, SARA-RT, and RT-Trajectory, are designed to enhance the speed, efficiency, and adaptability of robots in performing complex tasks in real-world environments.
AutoRT is a novel system that utilizes large foundation models to improve the training of robots. It combines these models with a robot control model to enable robots to collect training data in various environments. The system can direct multiple robots simultaneously, each equipped with video cameras and end effectors, to perform diverse tasks. In extensive real-world tests, AutoRT successfully managed up to 52 unique robots, gathering a dataset of 77,000 robotic trials.
Safety is a paramount concern in the integration of robots into daily life. AutoRT incorporates layered safety protocols, including a Robot Constitution inspired by Isaac Asimov’s Three Laws of Robotics and practical measures from classical robotics. These ensure that the robots operate within safe parameters and are constantly monitored by human supervisors.
SARA-RT, or Self-Adaptive Robust Attention for Robotics Transformers, is a new system that refines Robotics Transformer models, making them more efficient without compromising quality. It uses a method called “up-training” to reduce computational demands and improve the speed and accuracy of robotic decisions. The SARA-RT models have shown to be 10.6% more accurate and 14% faster than their predecessors.
RT-Trajectory addresses the challenge of translating abstract instructions into precise physical motions. It overlays visual outlines on training videos to guide the robot’s movements. This approach has significantly improved performance, with an arm controlled by RT-Trajectory doubling the task success rate compared to existing models.
These advancements are part of Google’s broader effort to build more capable and efficient robots. By integrating these systems, Google envisions a future where robots can perform a wide range of tasks with greater autonomy and effectiveness. The team is committed to ongoing research and adaptation to meet the evolving challenges and opportunities in robotics.