Three things I read this week on Physical AI, with a short note on each and a link to the source. The notes are mine. The reporting is theirs, so go and read it.

A robot model update, tested on a car production line

Google DeepMind introduced Gemini Robotics 2.5, an updated vision-language-action model, evaluated on more than 200 tasks including tool use, flexible materials and precision assembly. In a proof of concept at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik’s Apollo 2 humanoid used it for kitting: picking parts out of bulk containers and arranging them in trays, with a reported 96 percent success rate. The model is being made available to the company’s robotics partners. Robotics & Automation News

My note: Kitting is a good test because a bulk bin is messy. Every grasp starts from a slightly different pile. Two models split the work here: one plans the task, the other moves the arms and hands. Brains make a similar split between deciding what to do and working out how to move.

Warehouse robots that adjust, and software that coordinates them

DHL Supply Chain says it is testing adaptive robots and, after a proof of concept, has begun introducing AI orchestration software into operating warehouses. One application under test links autonomous mobile robots, packaging equipment and warehouse management systems, and shifts resources as order volumes change. The company reports more than 9,000 robots installed across its network and more than 1.5 billion robotic picks. Robotics & Automation News

My note: The line DHL draws is between physical AI, which moves material, and digital AI, which moves information. Definitions matter here because the words get used loosely. I sort out three neighbouring terms in Physical AI vs Embodied AI vs Robotics.

One million sidewalk deliveries

Avride says its delivery robots have completed one million deliveries in the United States, including 180,000 in October, and now run about 7,000 a day. It operates in more than 30 markets, among them Austin, Dallas, Jersey City, Philadelphia and Miami, with orders placed through apps such as Uber Eats and Grubhub. Robotics & Automation News

My note: A sidewalk is an unstructured place: curbs, dogs, rain, people looking at their phones. Volume matters to a machine that works there, because rare situations only show up when you do something many times. For more figures with sources, see Physical AI in Numbers.

More: all Physical AI news notes · what Physical AI is.

By Max Versace

Max Versace is an AI executive, scientist and entrepreneur specializing in brain-inspired AI, neuromorphic computing, Edge AI and Physical AI. He is VP of Emergent AI at Analog Devices and former co-founder and CEO of Neurala.