Peter Chen, Co-founder and CEO of Covariant, discusses the integration of AI in robotics and the challenges it faces. They explore the limitations of traditional robots and the role of AI in solving complex problems. The podcast also touches on building robot arms for business, the significance of ImageNet, robots in warehouse operations, AI models for robot arms, optimizing kitchens, GPT's impact, and transitioning from academia to industry.
Covariant aims to make AI breakthroughs work in robots by building a foundation model for robotics, similar to how GPT is a foundation model for language.
Covariant's model uses multiple cameras and an understanding of physics to enable robots to interact with heterogeneous objects in a warehouse, making it adaptable and reliable for industrial use.
Deep dives
Building AI Breakthroughs in Robotics
Covariant, a robotics company, aims to make AI breakthroughs work in robots. The co-founder and CEO, Peter Chin, explains that traditional robotics focuses on hardware and control algorithms. However, AI is needed to solve complex tasks that cannot be reduced to repeated motion, such as folding towels. Covariant's approach involves building a foundation model for robotics, similar to how GPT is a foundation model for language. By training the model on large robotics datasets, Covariant aims to create smarter and more capable robots for various industries.
The Power of Foundation Models
Covariant's foundation model is built specifically for robot arms used in warehouse and logistics. The model leverages multiple cameras to mimic human vision, allowing the robot to understand depth and make high-level inferences. Vision combined with an understanding of physics enables the robot to interact with heterogeneous objects in a warehouse, picking items with precision and efficiency. The model's power lies in its ability to generalize across different robot tasks, making it adaptable and reliable for industrial use.
Overcoming Data Constraints
Covariant recognizes the importance of large datasets for training AI models, but gathering robotics data poses challenges. Unlike language models that can scrape the internet, robotics data is limited. To address this, Covariant deploys its robot arms in warehouses to generate valuable data while performing real-world tasks. This data is used to improve the foundation model for robotics, making it smarter and more capable. Covariant's focus on building robot systems that work at scale and collecting data sets them apart in the AI robotics field.
Future Implications and Safety Considerations
Covariant's immediate focus is on safe, confined industrial robot applications. These robots follow strict design guidelines and compliance requirements, limiting safety concerns. However, as the technology advances and is applied to more free-form robots, safety considerations will become crucial. Covariant acknowledges the importance of addressing safety issues, but emphasizes that in the near term, their robot solutions are safe by construction, serving customers in warehouse and logistics sectors.
Peter Chen is the co-founder and CEO of Covariant. Peter’s problem is this: How do you take the AI breakthroughs of the past decade or so, and make them work in robots? Peter was one of the first employees at OpenAI, the maker of ChatGPT. On the show, he talks about how AI has evolved, and why it's so difficult to teach a robot to fold a towel.