Nebius and Positronic launch international robot manipulation AI competition

Nebius and Positronic Robotics have launched the international robot manipulation AI competition “Nebius × Positronic Physical AI Competition.” The primary source gives Oct. 26, 2026, as the announcement date. According to an official Nebius blog post dated Oct. 30, teams will compete in simulation, and the top 10 will be evaluated on physical robot arms in December 2026. Prizes include more than $20,000 in Nebius GPU credits and Positronic evaluation credits.

Positronic’s physical evaluations reportedly show a gap in average success rates between public models, from about 70% to about 12%, underscoring the need to compare robot manipulation models on the same hardware and tasks. By selecting teams through simulation before testing them on robot arms, the competition offers teams without physical robots a concrete route to participate in model evaluation.

The competition’s defining feature is its design of participation and evaluation, rather than the prize money. Teams do not need to own physical robots: they can submit models trained on Nebius AI Cloud, and those that rank highly can undergo Positronic’s physical evaluation. The structure provides separate access to GPU computing and evaluation using robot arms, limiting the latter to the top 10 teams. This can be read as an effort to separate access to physical validation from whether participants own their own robot equipment. The shared evaluation conditions also matter. The official blog says PhAIL.ai’s DROID evaluation uses the same Franka arm and covers more than 100 tasks. Comparing demonstration videos collected with different robots or on different tasks makes it difficult to determine whether performance differences come from the models or the evaluation conditions. The framework, with common hardware, tasks and scoring, standardizes comparisons within the competition at least. However, whether simulation rankings hold up on physical robots cannot be determined until the top 10 models have been tested on them. For Nebius, the competition links stages from training and simulation evaluation through physical validation, rather than simply providing cloud computing. The approach builds on the developer cloud described in Nebius partners with NVIDIA to build cloud for robotics and physical AI. Unlike the training and evaluation loop covered in Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando build closed loop for robot training in returns processing, this competition is designed to compare models from multiple teams through a shared benchmark and an open contest. Questions to watch include how closely simulation rankings match physical results, which tasks cause failures during physical evaluation, and whether results under the shared conditions apply to other robots and tasks. The official blog says contact, friction and timing differ in real-world settings. When interpreting the final results, the focus will therefore be not only on overall rankings but also on whether task-level results and the gap between simulation and physical performance are disclosed.

Submitted models will be evaluated in simulation, and the deadline for final entries is Dec. 13. Physical evaluation will use a Franka arm and a Robotiq gripper, with Positronic conducting tests as an independent evaluator. Nebius says teams will train models on its AI Cloud and submit them as containers. According to the official blog, the public DROID dataset underlying the evaluation environment contains 76,000 demonstrations collected across 564 scenes. PhAIL.ai’s DROID physical evaluation uses the same Franka arm and runs more than 1,000 episodes across more than 100 everyday manipulation tasks. The blog says the average task success rate for public models ranges from about 70% for the top model to about 12% for the lowest-ranked one.