Stanford AI100 Panel: Robots Cap at 26%, Waymo Is a Warning

Stanford’s One Hundred Year Study on Artificial Intelligence, AI100, has published its 2026 Study Panel Report, and its list of open grand challenge problems includes Physical AI and Embodied Intelligence. On the BEHAVIOR-1K benchmark, which tests robots in simulated homes, the report says “none have achieved scores above 26%”.

It also points to autonomous vehicles as “a cautionary tale”: Waymo and others have logged millions of autonomous miles but still represent only a tiny fraction of total trips. Physical AI is one of six grand challenge areas the section names, by our count, from a panel of 12.

Business Pill · Physical AI

An evergreen Business Engineer explainer on physical AI. It teaches the general idea only and does not discuss the AI100 report.

What the AI100 Panel Says About Physical AI

The section opens on Moravec’s paradox: “tasks involving high-level reasoning such as playing chess proved relatively easy for AI, while tasks requiring low-level sensorimotor skills, such as folding a shirt, remain difficult”.

It credits recent progress to “foundation models for robotics”, naming Robotic Transformer (RT-2) and Open X-Embodiment, and says advances in simulation are helping to close the gap between virtual training and physical hardware.

Then the limit: “a large gap remains between demonstration and deployment”. Over half a million factory robots are installed every year, the report says, but they “rely on highly constrained environments”, and consumer uses are limited to devices like camera drones and floor cleaners.

The Numbers in the Section

China, the report says, deploys more industrial robots than the rest of the world combined and “produces about 80% of the world’s humanoid robots, a sector that currently accounts for just 2% of the market but is growing rapidly”.

The best BEHAVIOR-1K score sits at no more than 26%, at least 74 points short of a full score, by our arithmetic. The core difficulties the panel lists are “long-horizon planning under uncertainty, achieving dexterous manipulation of diverse objects, and establishing formal performance guarantees”.

Where the Money Is Going

The report follows a run of physical-AI funding and launches in the past few days. Path Robotics said it raised $300 million at a $2.45 billion valuation, as covered in our Path Robotics piece. Nvidia introduced GR00T 2, pretrained mostly on human video.

Google DeepMind reported a 96% pick-and-pack success rate for a humanoid in a Mercedes-Benz proof-of-concept, set out in our Gemini Robotics piece. That is a single factory task, not a home benchmark, so it does not compare with the 26% BEHAVIOR-1K ceiling.

The Structural Read

As we read it, the panel’s own framing is the gap, not the absence of progress: it credits foundation models for robotics and better simulation, then says a large gap remains between demonstration and deployment.

As we read the section, the China figures cut two ways: China produces about 80% of humanoid robots, the report says, but humanoids are just 2% of the market, so the leader holds most of a small and fast-growing sector.

Three Implications

FOR ROBOT MAKERS The panel names long-horizon planning, dexterous manipulation and formal performance guarantees as the open problems.

FOR INVESTORS Path Robotics said on 10 November it raised $300 million at a $2.45 billion valuation, as covered in our earlier piece; the panel counts the field among AI’s open grand challenges.

FOR THE BENCHMARK No robot above 26% on BEHAVIOR-1K, the report says; a 96% factory pick-and-pack result from a Mercedes-Benz proof-of-concept is a different, narrower test.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The framework argues: “In any product cycle, there are two curves to follow.” One is what the technology can do; the other is what users can actually do with it.

As we read it, the AI100 panel describes physical AI with both curves apart: demonstrations from robotics foundation models on one, deployment limited to constrained factories and simple consumer devices on the other. The framework predates the report and does not discuss it.

What Is Not Established

The report’s figures on China and humanoid share are given without the market definition behind them in the section we read, and the section we read does not name the robot or team holding the top BEHAVIOR-1K score. The report’s citation dates it November 2026.

The Bottom Line

Stanford’s AI100 2026 panel lists Physical AI and Embodied Intelligence among six grand challenge areas, by our count, saying “a large gap remains between demonstration and deployment”: no robot scores above 26% on BEHAVIOR-1K, and autonomous vehicles from Waymo and others remain a tiny fraction of total trips.