Physical Intelligence Raises $600M to Develop Universal AI Robot Brains

Jeff Bezos-backed Physical Intelligence raises $600M to improve its AI robot brains - SiliconANGLE

Jeff Bezos-backed Physical Intelligence raises $600M to improve its AI robot brains Robotics software startup Physical Intelligence has reportedly closed on a fresh, $600 million round of funding that brings its valuation to a lofty $5.6 billion. The company has not officially disclosed the latest round, but reports by The Information and Bloomberg cite undisclosed sources who are familiar with the matter as saying it's true. Those sources said CapitalG, an independent growth fund operated by Alphabet Inc., led the round, with existing investors Lux Capital, Thrive Capital and Amazon.com Inc. founder Jeff Bezos all returning to throw more capital at the startup. New investors Index Ventures and T. Rowe Price also participated in the round, the sources said. Founded in 2024 by former Google DeepMind researchers and academics from Stanford University and the University of California at Berkeley, Physical Intelligence is developing artificial intelligence algorithms that can act as the "brains" of robots. The company, which is led by Chief Executive Karol Hasuman, says it's building a "general intelligence" system that will eventually be able to power any kind of robot for any kind of application. Robotics AI is a completely different kettle of fish compared to AI for chatbots like ChatGPT. Because robots operate in the real world, they need to be able to learn not only from text, but also other modalities, especially visual data, so they can perceive the physical environment they operate in and make decisions and take actions based on those inputs. Physical Intelligence has not yet announced any product or service, but Bloomberg said the company has been testing its AI software on robotic arms, performing tasks such as folding clothes, assembling boxes and making coffee. Earlier this week, the company released a new vision AI model based on reinforcement learning - a technique that enables models to improve over time as they learn from their experiences. The startup has shared a video of its robotic arms in action on X, saying that the reinforcement learning technique has helped it to double its overall throughput, or the number of tasks it can complete per hour. In testing, Physical Intelligence said the machines put to work for three hours, and were able to complete each task in an average of three minutes. To develop its models, the startup has spent a lot of money on gathering real-world data that's used for training purposes. A year ago, when it raised $400 million in funding, the company said one of the biggest challenges it faces is getting enough large-scale multitask and multirobot data to train its models, and its experiments allow it to generate valuable data itself. The startup says existing robots tend to be inflexible, because they're only designed to perform one particular task in a specific environment. They can work around small changes in their environments but can't adapt easily to extremely messy or complex spaces such as homes or other real-world places. Physical Intelligence's goal is to change that, using general intelligence to enable robots to learn and become acclimatized to any environment and task. The startup is not alone in thinking that AI can be used to create much more capable robots. Google LLC is also focused on robotic intelligence through its subsidiary Intrinsic, which just yesterday announced its partnering with Foxconn on a smart factory initiative. Meta Platforms Inc. is also developing AI for robots, and this week announced its latest Segment Anything computer vision models. Amazon has similar ambitions, and regularly introduces more sophisticated, AI-enabled robots in its logistics and warehouse facilities. Similarly, there are dozens of startups that compete with Physical Intelligence, with the likes of Gecko Robotics Inc., Genesis AI, Cobionix Corp. and FieldAI Inc. some of the latest to attract investment capital in recent months.

Bezos-backed Physical Intelligence raises $600M at $5.6B valuation

Robotics software startup Physical Intelligence raised $600 million in funding, valuing the company at $5.6 billion. The round, led by Alphabet's CapitalG, includes investments from Lux Capital, Thrive Capital, Jeff Bezos, Index Ventures, and T. Rowe Price. Founded in 2024, the startup develops AI algorithms to serve as robot brains. Physical Intelligence originated from a team of former Google DeepMind researchers alongside academics from Stanford University and the University of California at Berkeley. This founding group brings expertise in artificial intelligence and robotics to address challenges in physical automation. The company operates under the leadership of Chief Executive Karol Hausman, who oversees the development of core technologies. Hausman's role involves directing efforts toward innovative solutions in robotic intelligence. The primary focus of Physical Intelligence centers on creating artificial intelligence algorithms designed specifically to function as the operational "brains" of robots. These algorithms aim to establish a general intelligence system capable of powering various types of robots across diverse applications. Such a system would allow robots to handle multiple tasks without being limited to predefined functions, marking a departure from traditional robotic programming. Unlike artificial intelligence developed for chatbots like ChatGPT, robotics AI requires adaptation to physical environments. Robots must process inputs beyond text, incorporating multiple data modalities, with a particular emphasis on visual information. This multimodal approach enables robots to perceive their surroundings accurately, interpret physical conditions, and execute decisions that involve movement and interaction. Visual data processing proves essential for robots to navigate and respond to dynamic real-world settings effectively. Physical Intelligence has not released any commercial product or service to date. However, the company conducts internal testing of its AI software using robotic arms. These tests involve practical tasks such as folding clothes, which demand precise manipulation of flexible materials; assembling boxes, requiring spatial awareness and sequential actions; and making coffee, which includes handling liquids and equipment. These experiments demonstrate the software's potential in everyday manipulation scenarios. Earlier this week, Physical Intelligence introduced a new vision AI model grounded in reinforcement learning principles. Reinforcement learning involves models that refine their performance iteratively through accumulated experiences, adjusting behaviors based on outcomes. This technique allows the AI to optimize actions over time, improving efficiency and accuracy in robotic operations without explicit programming for every scenario. The company shared a demonstration video on X showcasing robotic arms utilizing the reinforcement learning method. In the video, the approach results in doubling the overall throughput, defined as the number of tasks completed per hour. During a controlled three-hour testing period, the robotic arms finished each assigned task in an average of three minutes. This performance indicates substantial gains in operational speed and reliability. To advance its AI models, Physical Intelligence allocates significant resources to collecting real-world data for training purposes. Data acquisition forms a critical component of model development, providing the diverse inputs necessary for robust learning. One year prior, during its $400 million funding round, the company identified obtaining sufficient large-scale multitask and multirobot data as a primary challenge. To overcome this, Physical Intelligence conducts experiments that generate proprietary data sets, enabling self-sustained improvement in training resources. Current robotic systems exhibit limitations in flexibility, as they receive design specifications for single tasks within controlled environments. These robots manage minor environmental variations but struggle with adaptation to highly disorganized or intricate spaces, such as residential areas or unstructured outdoor settings. Physical Intelligence seeks to address these constraints by integrating general intelligence into its algorithms. The general intelligence framework pursued by Physical Intelligence enables robots to learn from exposures and acclimate to varied environments and tasks. This capability would allow robots to operate effectively in complex, real-world conditions, transitioning from rigid programming to more autonomous functionality. By fostering adaptability, the technology supports broader deployment in practical applications beyond industrial confines. Several established technology firms pursue similar advancements in robotic AI. Google, through its subsidiary Intrinsic, announced a partnership with Foxconn on a smart factory initiative the day before the funding news. This collaboration aims to integrate AI into manufacturing processes for enhanced automation. Meta Platforms develops AI tools for robotics, including its recent release of Segment Anything computer vision models, which facilitate object detection and segmentation in visual inputs. Amazon incorporates AI-enabled robots into its logistics and warehouse operations, regularly deploying upgraded systems for tasks like sorting and transportation. These implementations demonstrate AI's role in scaling efficiency within large-scale facilities. The investments and developments by these companies highlight ongoing industry efforts to elevate robotic capabilities through artificial intelligence. Numerous startups also compete in the robotic AI sector. Recent examples include Gecko Robotics, which focuses on inspection technologies; Genesis AI, emphasizing autonomous systems; Cobionix, specializing in drone-based solutions; and FieldAI, targeting agricultural automation. Each of these firms has secured investment capital in recent months, reflecting investor interest in innovative robotic applications.