Building a Large World Model—the ChatGPT for the physical world

Embodied AI

Our AI software will power the transition to automation for the outdoor world. We are building robots with an AI brain, doing for the physical world what ChatGPT did for language. Where GPT knows how to predict which word should come next, our Large World Model (LWM) knows which action should come next in the real world.

We use deep learning to solve autonomy, eliminating costly robotic stacks that need intricate mapping and rules. The result is a data-driven algorithmic solution that thinks like a brain, continually learning from experience to maintain outdoor spaces in any environment, even new places, without needing training. We call it ES1, our GPT for outdoor work.

Simulation to Reality Data Engine

Our robots upload real data at the end of the day, and our simulation engine runs different permutations of feedback loops at the scale of 10,000 robots, for reinforcement learning at scale—with rapid improvement. 

The data engine is the true AI moat

Embodied AI’s growth hinges on data and large-scale reinforcement learning (RL) with tens of thousands of robots. Ultimately any model is only as good as the data it is trained on—and robotics is seriously constrained by access to relevant robot POV data and RL at scale.

While Sim2Real and passive data collection is useful – nothing can substitute a physical deployment and RL of robots at scale. This is our moat, and our right to win.

Our approach combines simulation-based pre-training and real-world fine-tuning, backed by a scalable data engine. This is the core of Large World Models (LWMs) and the future of automation.

ES1 Applications

Since it thinks for itself, ES1 can be used for all repetitive outdoor maintenance tasks, from snow plowing, to blowing, aerating, spraying, and mowing.