Boston Dynamics gives Spot the gift of gab

Boston Dynamics just slapped a mustache, googly eyes, and a ChatGPT brain onto its Spot robot, turning the high-tech quadruped into a sarcastic, chatty tour guide.

In a recent project detailed by lead software engineer Matt Klingensmith, the robotics company hooked Spot up to a cocktail of off-the-shelf generative AI models. The goal was simple: test whether foundation models could handle real-time autonomy and decision-making in the physical world without retraining the robot from scratch.

Python prompts and puppet mouths

Instead of overcomplicating the stack, the team relied on clever software duct tape.

Hardware modifications were minimal: a 3D-printed vibration mount held a ReSpeaker V2 mic array on Spot's payload bay. To give Spot some personality, engineers attached costumes to the arm gripper and used audio low-pass filtering to open and close the gripper like a talking puppet mouth.

On the software side, an offboard PC passed sensor data and prompts to OpenAI's GPT-3.5 and later GPT-4. Taking cues from Microsoft research, the team formatted prompts like python code documentation, limiting Spot's available actions to simple commands like moving to a spot, saying a phrase, or asking a question.

To make Spot aware of its surroundings, camera feeds from the body and gripper were routed through the BLIP-2 vision-language model once a second. Questions from humans were transcribed via OpenAI's Whisper, while ElevenLabs generated the spoken responses.

Sarcastic tours and yoga robots

The setup produced some surprisingly smart emergent behaviors. When a visitor asked Spot who company founder Marc Raibert was, the model replied that it didn't know and suggested visiting the IT help desk, promptly walking the group over there. When asked about its parents, Spot navigated to a display of retired BigDog and Spot V1 prototypes, describing them as its elders.

Roleplay worked remarkably well. When prompted to be sarcastic, Spot stayed in character and even led guests on a hunt for Bigfoot across the facility.

Still, the physical reality check hit hard. Latency hovered around six seconds between question and answer. Spot was also prone to standard LLM hallucinations, at one point assuring visitors that Stretch, the company's box-moving warehouse robot, was actually built for yoga.

Why it matters

We think this experiment exposes an ingenious shortcut for embodied AI. Instead of spending years building monolithic robotics models from scratch, Boston Dynamics treated GPT-4 as an improvisation actor and high-level planner, delegating the physical execution to Spot's battle-tested locomotion SDK.

For developers and robotics teams, it proves that combining off-the-shelf multimodal APIs can unlock believable conversational agents right now, without complex physics training.

However, the trade-offs are glaring. A six-second lag and frequent hallucinations might be charming on a novelty tour, but they are deal-breakers in industrial inspection or security. Relying on cloud connectivity also means the robot freezes the moment Wi-Fi drops or OpenAI's servers stutter. It looks impressive on video, but practical field deployment will clearly demand local execution and far tighter guardrails.