Google AI Teaches Robotic Dog Conversational Commands

2023-08-31

Google researchers may have finally found a technical solution - to complete the challenging task of training stubborn dogs to obey commands. Of course, the prerequisite is that you are willing to choose a robotic dog.

A group of research scientists from Google's DeepMind team have designed a large language model called SayTap, which can translate various commands that humans may give to dogs into a format that a four-legged robotic dog can understand. This model enables the robotic dog to understand basic commands such as walking forward and backward, as well as more contextual and complex concepts such as chasing squirrels or running quickly on a hot surface.

To achieve all this, DeepMind researchers assigned an expected "foot contact pattern" to the quadruped robot, where 1 represents a foot on the ground and 0 represents a foot off the ground. Then, these inputs were mapped to various combinations of the four limbs, teaching the robotic dog how to walk, trot, or jump in place. The researchers stated that these combinations of foot contact patterns are "a new interface between natural language user commands and motion controllers." As shown in the figure below, a researcher instructs the robotic dog to "trot forward." The bottom box shows a computer model translating this English command into task-specific binary code, which is then sent to the robotic dog, which dutifully executes it like an obedient child.

The researchers pointed out that successfully executing basic commands such as "forward" is not too complicated. What is more interesting is that they noted SayTap's ability to "handle unstructured and ambiguous commands." By providing the model with short prompts, the researchers successfully commanded the robotic dog to jump up and down when told "we are going on a picnic." In another test, when the operator told the robotic dog "the ground is very hot," it knew to run quickly. In perhaps the most amusing example, the robotic dog even backed away slowly when instructed to stay away from squirrels. Many real dog owners would long for this level of obedience.

"SayTap introduces an expected foot contact pattern as a new interface between natural language and low-level controllers," the researchers wrote in a recent blog post. "This new interface is simple, user-friendly, flexible, and allows the robot to follow direct commands and commands that do not explicitly specify how the robot should respond."

Although the general public now commonly associates large language models with their ability to generate book chapters or unique and bizarre images within seconds, StayTap's research highlights these systems' ability to transform complex human language into machine-understandable prompts. The researchers describe StayTap as a "controller that can be transferred to real robot hardware," but it is easy to see that similar advancements could one day enable people to truly interact with advanced robots simply by speaking to them, just like communicating with another human.

RECENT AI NEWS

RECENT AI TOOLS