Robot Rundown - 6/20/24

1. Ray Kurzweil on how AI will transform the physical world

Ray Kurzweil, a leading researcher in artificial intelligence over the last six decades, believes that AI is about to take a giant leap forward — moving from simply being digital tools to transforming the physical world. This article explains how AI will unlock enormous value across energy, manufacturing, and medicine.

• Energy: AI's ability to rapidly analyze billions of chemistries is revolutionizing the discovery of photovoltaic and battery materials, exemplified by Google's gnome AI increasing the number of known stable inorganic compounds from 20,000 to 421,000 overnight, paving the way for more efficient and cost-effective energy solutions.
• Manufacturing: AI-driven advancements in robotics and materials science are set to drastically reduce labor and raw-material costs, making goods cheap and abundant, while maintaining the exponential improvement in computing power. Fun fact: we have seen a 75 quadrillion-fold improvement in price-performance of computing since 1939.
• Healthcare: Up until 2022, we had determined the shape of approximately 190,000 proteins. That year, DeepMind’s AlphaFold 2 discovered over 200 million proteins, which will enable personalized treatments and potentially cure diseases like cancer and Alzheimer's.

Major Takeaway: Over the coming years, we will see AI’s impact transition beyond the digital world and into the physical world, drastically improving our outcomes in energy, manufacturing, healthcare, and beyond.

2. OpenVLA is an open-source generalist robotics model

You’ve heard of Large-Language-Models (LLMs), and you might have heard of Vision-Language-Models (VLMs) — but have you heard of Vision-Language-Action (VLA) models?

Researchers from Stanford, UC Berkely, Toyota, and Google Deepmind have introduced an open-source VLA model trained on a variety of real-world robotic scenarios. The researchers claim it outperforms similar robotic models, while leveraging optimization techniques that can run on consumer-grade GPUs for a low cost.

OpenVLA is a 7B-parameter model that can take natural language instruction and then perform actions to accomplish the desired task based on visual input. From here, there’s still plenty of work to be done to support multiple images and proprioceptive inputs.

Major Takeaway: OpenVLA democratizes access to advanced robotics technology, potentially leading to a surge in innovative applications and solutions in the robotics industry.

3. Towards Generalizable Embodied AI: Bridging Simulation and Reality

While most AI development has been done in conversational AI, such as ChatGPT, embodied AI focuses on physical embodiments of AI and tangibly interacting with the environment. Within the realm of embodied AI, vision-language-action models (VLAs) are becoming increasingly popular — these models handle multiple inputs, such as vision, language, and action modalities.

Historically, robots have primarily used reinforcement learning to focus on specific tasks, but there is a growing appetite for more versatile capabilities. Recent research shows promise for robots to complete complex tasks in diverse conditions through the use of VLAs. While these models promise a bright future, there are many challenges to overcome, such as scarcity of robotic data, real-time responsiveness, integration of multiple modalities, etc.

Major Takeaway: Bridging the simulation-to-reality gap is essential for advancing embodied AI, with significant implications for building robots that provide practical commercial value, allowing for a scalable mechanism to collect robot data and improve the performance of VLAs.