Obvious Ventures - Getting Physical: The New AI Frontier of Robotics - May 2024
FreeThe new AI frontier: bridging digital intelligence with physical robotics
About Obvious Ventures - Getting Physical: The New AI Frontier of Robotics - May 2024
This article from Obvious Ventures explores the intersection of generative AI and robotics, arguing that robotics represent the next frontier for AI to interact with the physical world. It discusses the limitations of current robotic systems—such as the fragility of pre-transformer neural networks and the high cost of tightly integrated hardware-software stacks—and highlights emerging breakthroughs including transformer-based architectures (e.g., DeepMind's RT-2), physics-based simulations (Nvidia's Isaac Gym), and advanced learning techniques like imitation learning and behavioral cloning. The piece includes a market map and positions companies like Figure, OpenAI, and DeepMind as key players in creating versatile, task-agnostic robotic brains.
Key Features
Pros & Cons
- Bridge between generative AI and physical world interaction
- Ability to learn new tasks with limited training samples
- Use of transformer-based models enables more robust and adaptable robots
- Simulation environments reduce need for expensive real-world training
- Pre-transformer neural networks require tens of thousands of training samples for simple tasks
- Tightly integrated hardware and software increases cost and limits flexibility
- High-performing systems like Boston Dynamics robots remain very expensive
- Current systems can be fragile when encountering novel objects or environments