This startup thinks robotics is about to have its ChatGPT moment
Summary
General Intuition, led by CEO Pim de Witte, is developing foundation models for embodied AI, aiming to replicate the success of large language models in natural language processing. The company argues that instead of extensive real-world data collection for specialized robot models, the industry should focus on high-quality datasets to create general-purpose models capable of transferring intuition across diverse environments. General Intuition built its foundation model by training on millions of hours of video game data, including human controller inputs, which de Witte and investor Vinod Khosla believe is key to spatial-temporal reasoning. The startup recently raised \$320 million at a \$2.3 billion valuation, demonstrating its model's ability to play video games and power a quadrupedal robot after fine-tuning with just eight minutes of real-world data. General Intuition's goal is to become the foundational AI for other robotics companies.
Key takeaway
For Robotics Engineers developing new autonomous systems, General Intuition's approach suggests a shift from extensive real-world data collection. You should explore leveraging foundation models trained on synthetic data, like video game interactions, to accelerate development. This strategy could drastically reduce the time and resources needed for fine-tuning, allowing your team to deploy capable robots with only minutes of task-specific real-world data. Consider integrating pre-trained general models into your development pipeline.
Key insights
Video game data can train embodied AI foundation models, enabling rapid transfer learning for diverse robotic tasks.
Principles
- General models reduce specialized data needs.
- Quality data trumps quantity for intuition.
- Spatial-temporal reasoning is key for embodied AI.
Method
General Intuition trained its foundation model on millions of hours of video game data, including human controller inputs, then fine-tuned with minimal real-world data.
In practice
- Develop general models for broad robotics applications.
- Explore synthetic data from games for robot training.
- Reduce real-world data collection for new tasks.
Topics
- Embodied AI
- Foundation Models
- Robotics
- Video Game Data
- Spatial-Temporal Reasoning
- General Intuition
Best for: Investor, Research Scientist, Entrepreneur, AI Scientist, Robotics Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Robotics News | TechCrunch.