Four 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.
Humanoids for the power grid
Neura Robotics, E.ON and the German energy company enviaM announced a partnership to bring robots to electricity networks. The robots are meant to inspect and monitor infrastructure such as substations and to help with heavy, repetitive or hazardous tasks. Training will happen at a robot gym in Munich, with first real-world trials planned for 2027. Robotics & Automation News
My note: A substation is a sensible first job for a robot: the place is known, the work repeats, and nobody enjoys it in January. What I will watch is how much of the training in the gym survives contact with a real site. That is the Sim2Real problem wearing a hard hat.
A factory robot that learns from the people next to it
The Swiss startup Embodied AI introduced Air, a wheeled humanoid for factory work. It learns manual tasks from demonstrations, and the company says the move to working alone usually takes about a month. Its arms are tendon-driven and built to give way if they hit an obstacle or a person. The AI Insider
My note: A month from demonstration to working alone is the number to remember. Learning on the job, next to people, is where I think this field gets decided. I wrote about why that is hard in Continual learning for robots.
A competition that ends on a real robot arm
Positronic Robotics and Nebius are running an international competition for robot manipulation models through December. Entries are scored in simulation first, then the ten best teams are tested on a physical robot arm. Robotics & Automation News
My note: I like that the last round is on a real arm. Simulation scores are easy to love and hard to trust.
More simulated worlds for robots to practice in
Physicl raised a 10 million dollar seed round led by Inovia Capital. Its product offers thousands of 3D environments and object assets for building scenes that produce synthetic training data for robots. TechCrunch
My note: Robots are short of data in a way language models never were. Nobody wrote down a few trillion examples of how to pick up a cup. Better simulated worlds are one of the ways the field is trying to close that gap. See world model in the glossary.
More: all Physical AI news notes · what Physical AI is.