DeepMind's Endeavor to Equip Robots with Real-World Skills

DeepMind's latest research endeavors aim to revolutionize robotics by training robots to operate efficiently in the real world. Through innovative AI models and a unique 'robot constitution,' they strive to fast-forward the evolution of robotics technology.

AutoRT: Advanced AI for Practical Goals

A pioneering system named AutoRT is central to their efforts. It leverages the strength of artificial intelligence to endow robots with an understanding of human tasks. These robots use a blend of foundational large models, like big language models or visual language models (VLM), and specific robot control models. This multifaceted approach enables the robots to accumulate experience in novel settings while expertly handling numerous tasks.

Synchronizing a Robotic Workforce

Described by DeepMind, AutoRT can seamlessly command an ensemble of robots, each outfitted with a camera and a manipulator, to undertake a variety of tasks across multiple environments. During extensive testing, which lasted seven months, AutoRT managed up to 20 robots concurrently and involved a total of 52 distinct robots operating in office buildings. This process resulted in a rich dataset of 77,000 robotic trials covering 6,650 unique tasks.

'Robot Constitution' for Safety and Ethics

To blend these robots safely into the human domain, DeepMind integrates a 'robot constitution' within AutoRT, embedding foundational safety regulations. The set incorporates Isaac Asimov's famous Three Laws of Robotics, emphasizing the paramount rule of not harming humans. Additionally, the robots feature numerous safety precautions, such as automatic shutdowns upon excessive joint force and oversight from a human supervisor equipped with emergency deactivation capabilities.

SARA-RT: Accelerating Robot Transformer Models

DeepMind's second initiative, named SARA-RT, significantly boosts robotic transformer models' effectiveness. The SARA-RT-2 models have shown remarkable improvements in both accuracy and speed when they process a brief sequence of images. This method provides a scalable mechanism to enhance transformer technologies, potentially broadening their application in robotics and other domains without costly pre-training.

RT-Trajectory for Task Generalization

The third project, RT-Trajectory, is developed to foster more versatile robotics. By incorporating visual cues into training protocols, the system aids robots in understanding and learning control policies. A robotic arm governed by RT-Trajectory demonstrated a 63% success rate on tasks outside its training purview, a significant leap from the earlier RT-2 models.

In addition, Google has also showcased a system enabling robots to autonomously generate code, which further enhances their capability to respond and conduct tasks effectively based on given instructions.

DeepMind, robotics, AI