The technology company, Google, and its artificial intelligence division, DeepMind, have announced significant progress towards developing safe autonomous robots for human interaction. The presentation introduces a new “Robot Constitution” designed to guide robot behavior in real-world situations.
The initiative was introduced alongside 3 pivotal tools: AutoRT, SARA-RT, and RT-Trajectory, each designed to enhance real-world data collection for robots, boost their speed, and improve their generalization abilities.
Robot Constitution
The Robot Constitution emphasizes the principle that robots must not harm humans. It prohibits robots from performing tasks involving the participation of people, animals, sharp objects, or electrical devices. Practical safety measures are also established, such as automatic shutdown if the force in the robot joints exceeds certain thresholds and physical shut-off switches for human control.
This work is based on Robotics Transformers 2 (RT-2), initiated by DeepMind to help robots make faster decisions and better understand and navigate their environments.
Breaking Down the Robot Constitution’s 3 Tools
1. AutoRT
AutoRT uses Large Language Models (LLM) or Visual Language Models (VLM) along with a robot control model (RT-1 or RT-2). It focuses on experiential training data collection in novel and diverse environments, allowing the simultaneous guidance of multiple robots in real-world settings. In simpler terms, AutoRT acts as the training master for robots, teaching them how to perform everyday tasks by showing them images and instructing them in various situations.
Over a 7-month period, AutoRT underwent evaluations in real-world environments, specifically in various office buildings. The test involved the safe orchestration of up to 20 robots simultaneously and a total of 52 unique robots, equipped with video cameras and end effectors, conducting 77,000 robotic tests in 6,650 unique tasks.
2. SARA-RT
SARA-RT transforms Robotics Transformers (RT) models into more efficient versions to enhance the speed and accuracy of RT models. Think of SARA-RT as an improvement to the quick-thinking ability of robots. It helps them think faster and more accurately when making decisions. SARA-RT was tested against existing RT-2 models and proved to be 10.6% more accurate and 14% faster after receiving a brief history of images.
DeepMind explains that they designed the system to be user-friendly, expecting researchers and professionals to apply it in robotics and beyond. Since SARA provides a universal recipe for speeding up Transformers without the need for computationally expensive pre-training, it has the potential to significantly expand the use of Transformer technology.
3. RT-Trajectory
This tool centers on refining robots’ movement generalization abilities, aiding them in interpreting and executing specific tasks through visual affordance of movement trajectories that are designed to address different aspects of robot interaction in the real world. It’s like a dance teacher that adds visual “paintings” of the robot’s movements onto training videos. In this way, the robot not only learns what to do but also how it should move to do the task efficiently.
The system can also create affordances by observing human demonstrations of the desired tasks, even accepting hand-drawn sketches. And it can be adapted to different robot platforms with ease. This model led to a two-fold increase in the success rate of existing RT models. In comparison to RT-2, an RT-Trajectory-controlled arm achieved a 63% success rate versus RT-2’s 29%.
Google and DeepMind have made notable progress towards creating autonomous robots safe for interaction with humans. The “Robot Constitution” and the 3 tools presented are significant milestones in the journey towards the safe integration of robots into society.
