NEWS  AI  January 5, 2024  Reading time: 2 Minute(s)

Max (RS editor)

DeepMind's robotics team made a significant leap towards enhancing the capabilities of robots operating in real-world environments, unveiling three groundbreaking advancements. These innovations, including the AutoRT data gathering system, a "Robot Constitution", and novel technologies like SARA-RT and RT-Trajectory, are designed to enable robots to make faster, better, and safer decisions.

AutoRT: Navigating the Complexities of Real-world Environments

At the heart of DeepMind's breakthrough is the AutoRT data gathering system. Leveraging a visual language model (VLM) and a large language model (LLM), AutoRT equips robots with the ability to understand their surroundings, adapt to unfamiliar settings, and choose appropriate tasks. This system, deployed in a fleet of 53 robots over seven months in four different office buildings, marks a significant stride towards real-world applications.

One notable feature of AutoRT is the incorporation of a "Robot Constitution", drawing inspiration from Isaac Asimov's "Three Laws of Robotics". This set of safety-focused prompts guides the LLM to avoid tasks that involve humans, animals, sharp objects, or electrical appliances, ensuring a safer interaction with the environment.

Safety Measures and Deployment Insights

DeepMind has prioritized safety in the development of these robotic systems. The robots are programmed to automatically stop if the force on their joints exceeds a specified threshold. Additionally, a physical kill switch is provided for human operators to deactivate the robots in emergency situations. The deployment of AutoRT involved 77,000 trials, with robots controlled remotely, operating based on scripts, or autonomously utilizing Google's Robotic Transformer (RT-2) AI learning model.

Next-Gen Technologies: SARA-RT and RT-Trajectory

DeepMind has introduced two additional technologies, SARA-RT and RT-Trajectory, to further enhance the capabilities of robotic decision-making. SARA-RT is a neural network architecture designed to augment the accuracy and speed of the existing Robotic Transformer RT-2. On the other hand, RT-Trajectory incorporates 2D outlines to assist robots in performing specific physical tasks with improved precision, such as wiping down a table.

Looking Ahead: A Glimpse into the Future of Autonomous Robots

While fully autonomous robots capable of serving drinks and fluffing pillows may still be on the horizon, these advancements represent a crucial step forward. The integration of safety protocols, decision-making frameworks, and next-gen technologies positions DeepMind's robotics team at the forefront of shaping the future of robotics in practical and dynamic environments. As we move closer to a new era of robotics, the lessons learned from systems like AutoRT may pave the way for more sophisticated and capable robotic assistants.

 MEDIA CREDITS: DEEPMIND/GOOGLE