Google has created a "Constitution for Robots"

Three new products — AutoRT, SARA-RT, and RT-Trajectory — were presented by the robotics team at DeepMind. They contribute to improving the speed of decision-making by robots, as well as enhancing their efficiency and safety when performing tasks in human environments.

AutoRT combines large foundational models, such as LLMs or VLMs, with RT-1 or RT-2. AutoRT can simultaneously control multiple robots, each equipped with a video camera and an actuator, to perform diverse tasks under various conditions.

SARA-RT makes models more efficient by using a new method for fine-tuning models. This upskilling approach transforms complex questions into simpler ones, making their resolution much faster and easier.

The RT-Trajectory model automatically adds visual contours describing the robot's movements to training videos.

One of the key innovations in AutoRT is the introduction of a "Constitution for Robots" within the LLM block. This concept ensures safety by instructing the machine to avoid situations involving people, animals, sharp objects, and electrical appliances. Additional safety measures include programming robots to stop when exceeding a certain force threshold, as well as incorporating an additional physical emergency switch in their design.

Source: deepmind.google