Physical
Introduction
We envision a future where humanity is no longer constrained by physical barriers. Just as software commoditized information, robotics will commoditize the physical world, unlocking abundance in products and enabling humanity's next frontier: building a spacefaring civilization with increasingly capable autonomous robots.
Over the past decade, advances in artificial intelligence have shown the power of large neural networks to model complex data. We can now generate coherent language, natural speech, and realistic video. Yet applying these modalities to interact with and manipulate the real world through robotics remains the unsolved challenge.
Challenges in Robotics
The main obstacle is obtaining high-quality data to train robotic systems. Three major approaches exist:
1. Teleoperated demonstrations
Robots can replicate tasks with enough expert demonstrations, but data collection is costly, requiring skilled operators and specialized hardware.
2. Simulation
Simulated training offers scale, but fidelity to the real world is limited, and building high-quality environments and dense rewards is costly and largely task-specific, often requiring a new setup for each manipulation task.
3. Human video demonstrations
Easiest to collect, since humans need only perform tasks naturally. The challenge has been accurately mapping human movements to robot actions. Recent advances in computer vision now enable high-fidelity hand and body tracking, making this approach increasingly viable.
Our Approach
We aim to leverage these breakthroughs in vision and robotics to begin the path toward physical automation. Our system will collect human demonstrations from clients and provide robots capable of performing those tasks.
Our long-term vision is a world where robots can reason, communicate, and physically act like humans, bringing true physical abundance.
Join us in building the future of physical automation