📖 Read more: ChatGPT in Robots: The AI Brain Controlling Machines
🚀 Claude on Mars: The First AI-Planned Drive on Another Planet
On December 8 and 10, 2025 (Sol 1707 and Sol 1709), engineers at NASA's Jet Propulsion Laboratory (JPL) used Claude to plan waypoints for the Perseverance rover navigating a field of rocks on the Martian surface. This wasn't real-time remote control — it was path planning done in advance, in 10-meter segments.
Anthropic announced “Claude on Mars” on January 30, 2026. The route covered roughly 400 meters. Using Claude Code, the model analyzed overhead images of Mars with its vision capabilities, plotted a breadcrumb trail point by point, and wrote commands in Rover Markup Language (RML) — a specialized XML-based language originally developed for the Mars Exploration Rover missions.
The waypoints were validated through a simulation that modeled over 500,000 variables — rover positions, terrain slopes, potential hazards. JPL engineers needed only minor corrections (such as splitting a route segment where ground-level cameras revealed sand ripples). JPL's estimate: Claude will cut route-planning time in half.
"Autonomous AI systems could help probes explore ever more distant parts of the solar system."
— Anthropic, “Claude on Mars,” January 30, 2026🐾 Spot + ChatGPT: The Robot Dog Tour Guide
The most notable demo of an LLM paired with a robot dog didn't come from Anthropic — it came from Boston Dynamics. In October 2023, their engineers published a detailed blog post titled “Robots That Can Chat,” describing how they turned the quadruped robot Spot into an autonomous tour guide at their facilities in Waltham, Massachusetts.
How It Worked
The architecture was deceptively simple at its core:
ChatGPT (GPT-3.5/4)
The brain: receives state information and decides what Spot should do — where to go, what to say, what to ask.
BLIP-2 Vision
Spot sees through its gripper camera. BLIP-2 converts images into text descriptions that ChatGPT can understand.
Whisper + ElevenLabs
OpenAI Whisper handles speech-to-text. ElevenLabs gives Spot a voice with customizable personality traits.
Simple API
Three commands: go_to(), say(), ask(). ChatGPT receives a list of locations and decides which command to execute.
📖 Read more: Amazon Warehouse Robots: Are They Replacing Workers?
The most striking part was the emergent behavior: when someone asked “Who is Marc Raibert?” (Boston Dynamics' founder), Spot decided on its own to walk over to the IT help desk and ask. When asked about its “parents,” it navigated to where BigDog and Spot V1 are displayed. Nobody had programmed these decisions.
"A world in which robots can generally understand what you say and turn that into useful action is probably not that far off."
— Boston Dynamics, “Robots That Can Chat,” October 2023🖥️ Computer Use: Bridging Text and Action
How does a language model control a physical robot? The answer lies in a capability Anthropic calls Computer Use, released in public beta in October 2024 alongside the upgraded Claude 3.5 Sonnet.
The concept: Claude can look at a screen, move a cursor, click buttons, and type text — exactly the way a human uses a computer. It was trained to interpret screenshots and count pixels for precise cursor positioning. On the OSWorld benchmark (screenshot-only), it scored 14.9% — nearly double the next-best AI system's 7.8%.
Why does this matter for robotics? Because it demonstrates that an LLM can execute the perception → reasoning → action loop — it sees visual information, decides what needs to happen, and performs precise actions. That's the same loop a robot needs. On Mars, Claude did exactly this: it looked at overhead images, reasoned about navigation waypoints, and wrote control commands.
📚 Timeline: LLMs in Robotics
🔬 How It Works: LLMs as Robot Brains
The fundamental idea behind every “LLM + robot” project is the same: separate thinking from doing. The LLM doesn't directly control motors or servos. Instead, it operates on three levels:
📖 Read more: Educational Robots: STEM Teaching Machines for Every Class
- Perception: Receives data — camera feeds, sensor readings, or text-based state descriptions
- Planning: Decides what needs to happen — which task, where to go, how to handle an object
- Execution: Produces commands in a language the robot understands — Python code, ROS commands, RML, or API calls
In practice, a safety layer sits between planning and execution — human verification, physics simulation, or a Robot Constitution (as in AutoRT). No serious project lets an LLM send raw commands directly to actuators.
Why Simulation Always Comes First
LLMs hallucinate — they produce plausible but incorrect outputs. In a chatbot, a hallucination is annoying. In a robot moving through physical space, it can be catastrophic. That's why on Mars, Claude's waypoints went through a simulation of 500,000+ variables before reaching Perseverance.
📊 Anthropic Today
Anthropic was founded in 2021 by Dario and Daniela Amodei, former OpenAI executives. As of February 2026:
Its major backers — Amazon ($8B), Google ($2B+), NVIDIA & Microsoft ($15B) — signal that Anthropic is far more than a chatbot company. Claude Opus 4.6 (February 2026) supports a 1-million-token context window, 128,000 output tokens, and leads benchmarks in agentic coding, computer use, and tool use.
🔮 Where This Is Headed
The trend line is unmistakable: LLMs are migrating from the digital world into the physical one. Today, Claude drives a rover on Mars. ChatGPT makes Spot give tours. Google's Vision-Language-Action models control dozens of robots at once. Every one of these achievements was unthinkable three years ago.
Anthropic hasn't (yet) announced an official partnership with a quadruped robot maker — but the technology is already there. The Unitree Go2's ROS 2 API ($1,600), Boston Dynamics' SDKs, Xiaomi CyberDog's open architecture — all that hardware is already waiting for a brain. The question isn't whether we'll see Claude or a comparable LLM controlling a robot dog in an actual home — it's when.
LLM-Powered Robotics: Who's Doing What
| Company / Lab | Project | Robot | LLM |
|---|---|---|---|
| Anthropic / NASA JPL | Claude on Mars | Perseverance Rover | Claude (Code) |
| Boston Dynamics | Robots That Can Chat | Spot (quadruped) | GPT-3.5/4 |
| Google DeepMind | RT-2 / AutoRT | Everyday Robots | PaLM-E, VLA |
| Figure AI + OpenAI | Figure 02 | Humanoid | GPT-4V |
| Microsoft Research | ChatGPT for Robotics | Various (framework) | ChatGPT |
What makes LLMs uniquely useful in robotics isn't that they “know” how a robot moves — it's that they can generalize. A traditional robotics system must be trained on every new task. An LLM can transfer knowledge from text, images, and experience to entirely new scenarios — exactly as Google's RT-2 achieved 62% success on tasks it had never seen before.
With Claude on Mars, Anthropic proved you don't need a specialized robotics model — a general-purpose LLM can learn an entire robotic language (RML) and write correct control commands. That same capability applies equally well to quadruped SDKs, robotic arms, or drones. The brain is ready. It's waiting for a body.