Tomorrow, Tested: Physical AI in Humanoid Robotics – A 2026 Landscape

A deep dive into the claims and evidence surrounding physical AI in humanoid robotics as of September 2026.

By Tomorrow, Tested.28 September 2026Updated: 28 September 2026Review responsibility: Tomorrow, Tested.

Tomorrow, Tested.

EDITORIAL CARD

This article explores the current state of physical AI in humanoid robotics, focusing on claims made by companies like Unitree, NVIDIA, and BMW, while emphasizing the need for independent verification.

What we know

Unitree Robotics claims UnifoLM-X2-1.0 enables fully autonomous combat for humanoid robots: Unitree Robotics' video from September 7, 2026, states that UnifoLM-X2-1.0 'breaks through world-action foundation models’ bottlenecks' in planning, decision-making, and dynamic execution, achieving 'fully autonomous combat' for humanoid robots. This validates the fundamental feasibility of large-scale deployment of world model-driven humanoid robots. ↗
Unknown
XPENG activates production line for IRON humanoid robot: XPENG announced on September 8, 2026, that its first advanced humanoid robot, IRON, has walked off the production line, marking a key step toward mass production by late 2026. ↗
Unknown
BMW Group deploys physical AI in Leipzig Plant pilot project: BMW Group Plant Leipzig is running a pilot project that brings physical AI to Europe, with the new competence centre driving the integration of artificial intelligence (AI) and robotics worldwide. The project includes tasks like lifting heavy components. ↗
Unknown
NVIDIA announces physical AI tools for robotaxi and humanoid applications: NVIDIA announced on September 12, 2026, that it is developing physical AI for robotaxi and humanoid applications, including tools like Jetson Thor robotics, Cosmos world model, and DRIVE Hyperion 9. ↗
Unknown
Agility Robotics unveils Digit 5 humanoid robot: Agility Robotics unveiled Digit 5, the next generation of its general-purpose humanoid robot, on September 15, 2026. ↗
Unknown
Renesas establishes Physical AI & Robotics Lab in Beijing: Renesas announced on August 26, 2026, the official opening of its Physical AI & Robotics Lab in Beijing, China. ↗
Unknown

The 2026 landscape of physical AI in humanoid robotics is marked by ambitious claims and limited transparency. Unitree Robotics’ video from September 7, 2026, asserts that its UnifoLM-X2-1.0 model has overcome bottlenecks in planning and decision-making, enabling ‘fully autonomous combat’ for humanoid robots. While this validates the feasibility of world model-driven systems, no independent testing or benchmarks are provided to substantiate these claims.

NVIDIA’s announcement of physical AI for robotaxi and Skild AI, paired with tools like Jetson Thor and DRIVE Hyperion 9, highlights a focus on industrial and transportation applications. However, the description lacks specifics on how these systems perform in real-world scenarios. Similarly, OpenAI’s 2027 demo of a humanoid robot is described as ‘impressive,’ but without technical details or test outcomes, it remains speculative.

Industry adoption is evident in projects like BMW’s pilot at Leipzig Plant Leipzig, where ‘physical AI’ is used for tasks such as lifting heavy components. Yet, the description does not clarify the system’s design, safety protocols, or scalability. XPENG’s activation of a production line for a general-purpose humanoid robot and Agility Robotics’ Digit 5 reveal progress in manufacturing, but neither provides evidence of real-world performance or customer adoption.

The absence of independent validation raises questions about the practicality and reliability of these systems. While the claims suggest significant advancements, the lack of concrete data underscores the need for further scrutiny and transparency in the field.

Tomorrow, Tested: Physical AI in Humanoid Robotics – A 2026 Landscape

What we know

Sources & provenance