Robot brain builders are pushing out of their GPT-2 era

The venture capital landscape is currently obsessed with "Physical AI," a sector where billions of dollars are being poured into applying the sophisticated architectures behind large language models (LLMs) to the world of robotics. This fervor recently culminated in a massive IPO for Unitree, China’s premier robotics firm, which saw its valuation soar to $66 billion upon hitting the public markets. However, the market’s enthusiasm hit a harsh reality check this week, as the company’s valuation plummeted by nearly 50%.

Industry analysts point to a fundamental disconnect: while the mechanical prowess of these robots is advancing rapidly, their "brains" lack the cognitive sophistication required to perform tasks that generate genuine economic value.

The Robotics Data Crisis

The tension between hype and utility was palpable at last week’s Actuate conference, a premier gathering for developers focused on building AI brains for robots. The event has seen explosive growth, tripling in size since its 2023 debut to host 1,500 attendees. The conference was organized by Foxglove, a firm specializing in data visualization and management for physical AI.

Despite the energy, the industry is grappling with a significant bottleneck. A booth for Avala, an infrastructure startup, displayed a sign promising to solve the "robotics data crisis." This crisis stems from a severe shortage of high-quality, diverse training data necessary to make AI models truly capable in the physical world.

"Physical AI is in its 'GPT-2 era,' the OpenAI model that pre-dated the arrival of ChatGPT. More data and compute will be needed to get over the hump, particularly GPUs optimized for ray tracing, which are used to create high-fidelity simulations." — Harry Mellsop, founder of Antioch.

The Path to Generalization

Currently, the dream of a "general-purpose" robot—one capable of performing any task—remains a distant horizon. End-to-end learning models have yet to deliver products that offer the reliability required for commercial deployment. To bridge this gap, developers are looking to emulate the strategies of frontier AI labs: diversifying datasets, experimenting with novel training regimes, and refining reinforcement learning scenarios.

Autonomous vehicles (AVs) currently lead the pack in this transition. Their advantage is twofold: they benefit from massive amounts of real-world data collected from human drivers, and their primary objective—avoiding collisions—is fundamentally simpler than the complex physical manipulation required of humanoid robots.

Many of the tools currently powering the physical AI revolution were born in the AV sector. Foxglove, for instance, was founded by veterans of Cruise, the self-driving unit formerly under General Motors. Now, automotive giants and rideshare platforms are pivoting, betting that their existing machine learning infrastructure will allow them to outpace dedicated humanoid robotics startups. Tesla is aggressively pursuing this with its Optimus program, while Wayve and Uber have launched dedicated R&D labs to explore humanoid form factors.

Hardware vs. Software: The Great Debate

The industry is currently split on the best strategy for development. Alex Kendall, CEO of Wayve, advocates for a hardware-agnostic approach.

• The Case for Agnosticism: Kendall believes it is premature to commit to specific hardware platforms, as sensor and component technology is evolving too quickly.
• The Case for Vertical Integration: Théophile Gervet, CEO of Genesis AI—which recently raised a $105 million seed round—argues that we are in the early stages where co-designing hardware and AI is essential for success.

Gervet highlights a critical dilemma: general-purpose humanoids are currently stuck in the lab, while task-specific robots are already generating revenue in the field. Companies like Gritt (solar farms), Agility (industrial settings), and Bedrock (autonomous excavators) are finding success by focusing on narrow verticals.

"No customer cares about the general-purpose robot that works at 80% success rate. We see a lot of other players go general, but there is no value provided because there’s no vertical focus. But then, if you’re building for a narrow vertical on top of GPT-2, you’re going to get crushed by the company building on GPT-4." — Théophile Gervet, CEO of Genesis AI.

The Quest for the 'ChatGPT Moment'

The industry is collectively searching for its "ChatGPT moment"—a breakthrough that will signal the arrival of reliable, consumer-ready physical AI.

For Alex Kendall, that moment isn't about investor sentiment, but consumer impact. He envisions a scenario where "eyes-off" autonomy becomes available for less than $1,000 in hardware costs. By licensing models to car manufacturers, Wayve aims to build a foundation for a truly general embodied AI.

Théophile Gervet defines the milestone differently: "Manipulation that just works out of the box. You can talk to a robot in natural language and have it do any basic task for manipulation... and it works to some level of reliability, let’s say 80% plus out of the box—that’s roughly your ChatGPT experience."

However, Adrian Macneil, CEO of Foxglove, offers a more grounded perspective. He argues that the "ChatGPT moment" was defined by its massive, rapid distribution—a feat that is significantly harder to replicate in the physical world where hardware deployment is a major hurdle.

Key Takeaways for the Future of Robotics

• Data Infrastructure: The industry is shifting focus toward better data management. Foxglove’s new product, built on Nvidia’s Cosmos open-weight model, allows engineers to use natural language queries to debug and evaluate robot performance.
• Simulation is Key: High-fidelity simulations are the primary training ground for these AI brains, necessitating a massive increase in compute power and specialized GPU usage.
• The "Apple II" Goal: Rather than a singular viral moment, the industry may be heading toward an "Apple II" or "IBM PC" moment—a time when home robots become accessible, useful, and fun for the average consumer.

As the sector moves out of its "GPT-2" phase, the winners will likely be those who can balance the need for high-quality, vertical-specific data with the long-term goal of creating a truly general, reliable, and affordable embodied intelligence.