The realm of physical artificial intelligence is a vibrant and expanding domain within venture capital, drawing substantial investments aimed at integrating large language model-like intelligence into robotic systems. This burgeoning interest highlights a critical juncture for the industry.
The Evolving Landscape of Robot Intelligence: From GPT-2 to Future Horizons
In May 2026, Tokyo witnessed the Humanoids Summit, where Kiyoshi Ota/Bloomberg captured a significant moment: an exhibitor demonstrating a Unitree Robotics G1 humanoid robot. This image encapsulates the current state where robotic physical prowess often outstrips their AI decision-making capabilities. Just last week, China's foremost robot manufacturer, Unitree, experienced a substantial market correction on its exchange debut, losing nearly half its valuation. Analysts largely attribute this decline to a fundamental gap: while the robots exhibit impressive physical attributes, their underlying artificial intelligence has yet to mature sufficiently for complex, value-generating tasks.
The Actuate conference, a recent gathering of innovators focused on developing AI 'brains' for robots, illuminated both the enthusiasm and the hurdles within the sector. Organized by Foxglove, a firm specializing in data management and visualization for physical AI model development, the event has seen its attendance triple since its inception in 2023, attracting 1500 participants. The palpable excitement was, however, tempered by an underlying concern: the scarcity of high-quality training data for AI models. A booth for Avala, another key infrastructure provider in physical AI, prominently displayed a sign declaring a commitment to resolving the 'robotics data crisis.' The consensus among developers points to a need for more diverse datasets, innovative training methodologies, and improved reinforcement learning environments to emulate the successes seen in cutting-edge AI research.
Harry Mellsop, co-founder of Antioch, a startup focused on simulation tools for model builders, posited that physical AI is currently in its 'GPT-2 era,' drawing parallels to the early stages of OpenAI's large language models. He emphasized that overcoming current limitations will necessitate significant advancements in data collection and computational power, particularly GPUs optimized for ray tracing to facilitate high-fidelity simulations. Autonomous vehicles currently lead in this regard, benefiting from ample real-world driving data and a primary objective of collision avoidance rather than complex physical manipulation. Many of the tools used for building these models originated from autonomous vehicle companies; for instance, Foxglove was established by former employees of Cruise, General Motors' self-driving division.
Automotive companies, including Tesla with its Optimus robot, are increasingly leveraging their investments in machine learning tools to compete in the humanoid robot space. Even autonomous driving-focused Wayve and ride-sharing giant Uber have established robotics laboratories dedicated to humanoid research and development. Alex Kendall, CEO of Wayve, suggested that while core data and simulation infrastructures might be shared, the 'world models' for simulators would require distinct post-training for different robotic embodiments. He advocated for hardware agnosticism, given the rapid progress in sensor and component technology.
Conversely, Théophile Gervet, CEO of Genesis AI, a vertically integrated humanoid robotics company that secured a substantial seed funding round this year, argued that early stages of this technological wave favor a co-design approach for hardware and AI. Gervet also addressed the ongoing debate regarding the focus of physical AI businesses. Companies targeting specific tasks, such as Gritt in solar farm construction, Agility in industrial settings, and Bedrock in autonomous excavation, are successfully deploying robots, generating both revenue and invaluable real-world data. However, general-purpose humanoids continue to reside primarily within research labs. Gervet highlighted the challenge: customers demand high reliability (80% success rate or more) for practical applications, and a narrow vertical focus risks being outpaced by more advanced, general-purpose models. Kevin Peterson, CTO of Bedrock, explained that his company's initial focus on excavation serves as a strategic entry point to understand 'manipulation in the wild,' with aspirations to develop a broader intelligence layer for various construction machinery.
The management of vast quantities of data, especially visual and lidar information, presents a significant challenge. Foxglove recently unveiled a new product, built upon Nvidia's Cosmos open-weight world model, designed to enable engineers to search and analyze this data using natural language queries, thereby accelerating evaluation and simulation processes. This aims to speed up debugging and iteration cycles for model developers.
The elusive 'ChatGPT moment' for physical AI, which OpenAI's Sam Altman anticipates within a few years, remains a topic of discussion. Kendall believes this breakthrough would captivate consumers, rather than just investors, much like household vacuum robots. He envisions a future where 'eyes-off autonomy' is achievable in cars for less than $1000 in hardware costs, a goal his company actively pursues through licensing models to automakers. This, he argues, represents a multi-billion dollar opportunity that will pave the way for truly general embodied AI.
Gervet's vision for this transformative moment involves robots capable of executing 'manipulation that just works out of the box.' This would entail natural language interaction for basic tasks like pushing, pulling, closing laptops, or cleaning, with a high degree of reliability (over 80%).
Adrian Macneil, Foxglove's CEO, offered a different perspective, asserting that there might not be a single 'ChatGPT moment' for robotics due to the inherent complexities of real-world distribution. Instead, he expressed eagerness for an 'Apple II' or 'IBM PC' moment in robotics, signifying the widespread availability of useful and enjoyable home robots.
The future of physical AI hinges on bridging the gap between sophisticated hardware and equally advanced cognitive capabilities. As researchers and companies continue to innovate, the debate between specialized applications and general intelligence will undoubtedly shape the trajectory of this exciting field.
