How they use Nebius AI Cloud

Robot Company uses Nebius for robotics development experiments alongside local GPU work. Its soft-surface picking project trains a robot gripper to lift flat washers from compliant materials. Nebius Serverless AI Jobs run simulated dataset generation, evaluation, material and maneuver optimization, and policy training. A CPU Serverless Endpoint serves the trained grasp-success scorer over HTTPS. Joint material and movement optimization remains work in progress.

For a separate humanoid-motion exercise, Robot Company fine-tuned NVIDIA’s SONIC controller in Isaac Lab. A local RTX 5080 trained the reference motion while a Nebius H100 trained a more demanding variant using 4,096 simulated environments and 500 iterations. The company reports that this parallel setup halved its iteration loop. The cloud job assembled its environment from public images and retrieved a private checkpoint and motion file from object storage.