Figure01, Blackwell, and the Race to Create the First General Purpose Robot

April 7, 2024

Since the late 1960s, academic research in computer vision has led to advances in computer’s ability to process, analyze, and extract data from the real world through digital images. Not far behind came machine vision, the underlying technology used to teach computers to recognize objects and people in images and videos. From facial recognition in smartphones to security surveillance, computer vision is progressing at an impressive rate. In the modern day, computer vision systems such as that in Tesla’s suite of autonomous vehicles display an impressive understanding of the world around them.

Over the past few years we have seen the explosive growth of generative AI and machine learning, with the culmination of these two in ChatGPT in November of 2022. The release of this model opened the ‘Pandora’s Box’ of language models. Within months, we began to see new models from other companies such as Meta’s LLaMA, Anthropic’s Claude, Google Gemini, and many more. With the release of these foundation models, people began to realize that computers are now able to speak to us as would another human.

As the capabilities of machines increase, so does the feasibility of creating machines with the ability to speak, listen, and reason. Big players in the software and hardware industries are now racing to develop the first commercial general purpose robot. A robotics startup called Figure has recently partnered with OpenAI to pursue this development. Figure demonstrated their prototype model which was able to speak to a human, understand AND manipulate its environment upon request. It is able to achieve these tasks fully autonomously.

Despite the abilities of the Figure01 and other cutting edge prototypes such as NVIDIA’s Project GROOT, these robots still have some downsides. While these robots are able to accept and process instructions in natural language, they take several seconds to process them. The robots are also limited in their physical dexterity, and move quite slow compared to humans. The last bottleneck is the available computational power required to bring these robots up to speed.

NVIDIA just held their yearly keynote, in which they demonstrated their current research. The most impressive development was their NVIDIA GB200 NVL72 using the Blackwell architecture. This system is built for processing enormous AI models and has 11.5 exaflops of AI supercomputing at FP4 precision with 240 terabytes of fast memory. With NVIDIA leading the charge of computer hardware and Big Tech racing to develop AI models, prepare to see some remarkable (and potentially frightening) advances in AI ability. In the next few years, we will all be witness to the race to the first general purpose robot.