What happened
Nebius and Positronic Robotics have launched an international competition focused on advancing robotic manipulation models. The initiative, which runs through December 2026, invites research teams to submit AI models for evaluation. The process begins with simulation-based testing, after which the top 10 teams will have their models deployed on a physical Franka robotic arm to assess real-world performance. The competition is supported by over $20,000 in GPU computing and evaluation credits provided by the organizers.
Nebius and Positronic Robotics announced the launch of their joint physical AI competition on October 26, 2026. The event is designed to test the capabilities of robotic foundation models in performing everyday manipulation tasks.
The competition structure involves a multi-stage evaluation process. Initially, all submitted models undergo testing in a simulated environment. The 10 teams that demonstrate the highest performance in simulation will advance to the final stage, where their models will be tested on a physical Franka robotic arm.
The organizers are offering a prize pool exceeding $20,000, distributed as Nebius GPU computing credits and Positronic evaluation credits. This infrastructure support is intended to lower the barrier to entry for smaller research teams.
Nebius provides the underlying GPU infrastructure and its Physical AI Workbench, while Positronic manages the evaluation process through its PhAIL.ai leaderboard and the DROID .
Source details: roboticsandautomationnews.com
Why it matters
This competition addresses a critical bottleneck in the development of embodied AI: the lack of standardized, independent benchmarks for robotic foundation models. By moving beyond simulation to physical testing, the organizers aim to provide a clearer picture of how models perform in real-world manipulation tasks. The disparity in current model performance—ranging from 12% to 70% success rates in existing benchmarks—highlights the volatility of current research. Establishing a consistent evaluation framework, similar to the historical impact of the DARPA Grand Challenge, is essential for maturing the field and enabling smaller teams to contribute meaningful advancements to robotic autonomy.
The current landscape of robotic AI is fragmented, with few standardized methods for comparing how different foundation models handle physical tasks. Positronic reports that existing evaluations using the DROID show a massive performance gap, with top models succeeding in 70% of tasks compared to only 12% for the lowest-performing models.
By creating a standardized, independent testing environment, the organizers hope to replicate the success of the DARPA Grand Challenge, which historically accelerated the development of autonomous vehicles by providing a clear, competitive goalpost for researchers.
The competition serves as a practical test of whether current AI models can bridge the 'sim-to-real' gap, a persistent challenge in robotics where models trained in virtual environments fail to translate their skills to physical hardware due to sensor noise, friction, and other real-world complexities.
This initiative is a significant move for the industry, as it shifts the focus from theoretical model capabilities to verifiable, physical performance, which is a prerequisite for the widespread deployment of autonomous robots in industrial and domestic settings.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
The primary focus is the transition from simulation to physical hardware for the top 10 finalists. Observers should monitor whether the performance gap observed in simulation persists during physical execution, as this will reveal the of current foundation models in handling real-world environmental variables. Additionally, the competition's reliance on the DROID and the Franka robotic arm provides a specific, controlled environment that may influence which types of models succeed. Future updates will likely center on the final performance metrics and whether this initiative successfully establishes a new industry standard for physical AI evaluation.
The competition is scheduled to run through December 2026. The most significant milestone will be the announcement of the 10 finalists and their subsequent performance on the physical Franka robotic arm.
Industry observers should watch for whether the results of this competition lead to the adoption of the DROID as a broader industry standard for evaluating physical AI.
The organizers have not disclosed specific details regarding the long-term accessibility of the evaluation platform for researchers outside of the competition, which remains a meaningful unknown for the broader robotics community.
The competition's success will be measured by its ability to provide consistent, reproducible data that can help developers identify why certain models outperform others in physical manipulation tasks.