Bleu Robotics raises €5 million to teach factory humanoids a task from about 10 demonstrations

Bleu Robotics, a Paris startup that writes software for industrial humanoid robots, has raised a €5 million pre-seed round co-led by Contrarian Ventures and OVNI Capital. Kvanted, WEPA Ventures, Motier Ventures and Charlie Songhurst also invested. The founders are Inria research directors Jean-Baptiste Mouret and Serena Ivaldi, and Benoît Berkoukchi, who ran assembly-line teams at Renault.

Bleu bets that factory robots need less learned perception rather than more data. Its Bleu_0.1 policy is not pre-trained. Depth cameras and off-the-shelf trackers such as Meta's SAM2 and Nvidia's FoundationPose locate the objects, and the model learns only the task from their positions. Bleu says a robot can learn a task from fewer than 10 demonstrations in under an hour. In the ManiSkill3 benchmark, it reports a 98% success rate with fewer than 10 demonstrations, while diffusion policies needed at least a hundred.

London's Machenta, which raised a $40 million seed two days earlier, took the opposite route and pre-trains its model entirely in simulation. Bleu's route skips the pre-training bill, and Bleu itself names where it stops: fabric is very hard to track. The company's own imitation learning tests used rigid workshop items: loading a drilling station, and picking spray cans and glue bottles into a bin, on a Unitree G1 and a Franka arm.

Startup Rise says Bleu already works with two major industrial companies. The first one it names will show whether its parts can be tracked.