THE ESSENTIALS IN 30 SECONDS

  • Bleu Robotics, founded in Paris in December 2025, raised a €5 million pre-seed round co-led by Contrarian Ventures and OVNI Capital.
  • Its hardware-agnostic software infers the goal behind an operator's demonstration; the company says about ten demonstrations are enough to learn a task.
  • The startup says first factory deployments are under way and aims to cut a full humanoid deployment to under an hour; its accuracy claims have not been independently verified.

Bleu Robotics, a Paris startup that writes software for industrial humanoid robots, has raised €5 million in a pre-seed round announced on October 30. Contrarian Ventures and OVNI Capital co-led the round, with Kvanted, WEPA Ventures, Motier Ventures and technology investor Charlie Songhurst also taking part, according to Robotics & Automation News.

A research spin-off with a factory focus

The company was founded in December 2025 and grew out of Inria's HuCeBot humanoid robotics team in Nancy, French Tech Journal reports. Its chief executive, Benoît Berkoukchi, previously ran assembly-line teams at Renault. The two other co-founders, Jean-Baptiste Mouret and Serena Ivaldi, are Inria research directors with about 15 years of work in AI and robotics behind them. French Tech Journal puts the team at eight people.

Bleu does not build robots. It sells a learning layer meant to run on any humanoid platform, which French Tech Journal says works with robots connected through ROS and would be sold as a monthly subscription through robot makers and integrators.

Learning from a few demonstrations

Most robotics companies try to train general-purpose robots on very large datasets. Bleu takes a different route that it calls abductive learning: the model tries to infer the goal and constraints behind what a human operator shows it, then works out how to repeat the task reliably. Operators teach at the workstation with handheld grippers and no code, and the company says about ten demonstrations are enough for a new task.

Bleu claims accuracy above 99% on production tasks. The AI Insider notes that the announcement came without test methodology, so the figure cannot be checked. The one public benchmark cited by French Tech Journal is the company's first model, Bleu_0.1, which it says reached 98% success in March on a cube-stacking task from the ManiSkill3 simulation suite. French Tech Journal also lists limits the founders acknowledge: the system is not designed for soft objects such as cloth, and it asks for help when a situation falls outside what it has seen.

Customers and next steps

The company says it is working with two large industrial firms. According to The AI Insider, first deployments involve automotive manufacturers it does not name and WEPA, a European hygiene-paper producer whose venture arm is also an investor. The money will go into hiring and into extending the software to more industries, with a stated goal of bringing a complete humanoid deployment down to under an hour. Bleu is still pre-commercial, and these deployments remain early.