ChatGPT Gives Boston Dynamics Robot Dog Personalities

Boston Dynamics is using ChatGPT to give its robot dog Spot a range of personalities, as well as the ability to converse with people, explain its environment and adapt its behavior as it goes. 

Using the large language model, Spot can now understand and respond to questions, identify and explain objects in its environment, and autonomously plan actions in response to different scenarios.

The team said the project turns Spot into a “chat (ro)bot,” bringing the advancements from generative AI into dynamic robotic applications.

In a video demonstrating Spot’s new capabilities, the robot is shown giving visitors a tour of the company’s facilities and assuming a number of personalities including a British butler, a 1920s archaeologist, a teenager and a Shakespearean actor.   

A new “mouth” has been attached to Spot’s body which moves as it speaks, “like the mouth of a puppet,” the company said.

“Over the past year or two, advances in artificial intelligence, specifically … Generative AI have been rapid,” Boston Dynamics said on its website. “We wanted to explore how these models work and how they might impact robotics development.”

In particular, the company said it wanted to look at large Foundation Models (FM) – large AI systems trained on a vast dataset. 

These large models have a capability known as Emergent Behavior, meaning they can perform tasks beyond the scope of what they were programmed to do, enabling greater autonomous capabilities when deployed in robots. 

“Because of this,” the company said, “they can be adapted for a variety of applications, acting as a foundation for other algorithms.”

Using FM systems, Spot was able to make real-time decisions, and the team also leveraged Visual Question Answering (VQA) models which allow Spot to identify and “caption” images, before generating responses about them. 

“We gave the robot a very brief script … for instance a sentence that says ‘this is the charging station that's where robots gonna recharge,’” said Matt Klingensmith, Boston Dynamics’ principal software engineer. “And it combines that with imagery from its cameras and it runs it through what's called a Visual Question Answering model to try to get more information about what it sees before generating a response. 

“So literally, it will caption what's in the images around it, feed that in with the script that we gave it, and come up with a response.”

The team said there were some surprises along the way. In one example, when asked who its parents were, Spot took the team to older models of Spot on display, referring to them as its “elders.”

While the team said the behavior doesn’t suggest any level of consciousness or unexpected intelligence, it did demonstrate the “power of statistical association” and Spot’s capacity to connect concepts.

Boston Dynamics said it is continuing to explore the deployment of AI in robots, adding that a future where robots can understand and respond to humans is “not that far off.”

“Robots provide a fantastic way to “ground” large foundation models in the real world,” the team said. “By the same token, these models can help provide cultural context, general commonsense knowledge, and flexibility that could be useful for many robotics tasks … That kind of skill would enable robots to perform better when working with and around people – whether as a tool, a guide, a companion or an entertainer.”