Lumos Robotics says its Prime R0 industrial embodied AI model has achieved the highest overall score on the latest MolmoSpaces leaderboard, outperforming larger models from competitors including Nvidia and research teams from the United States.
The Chinese robotics company said its 2.8-billion-parameter model ranked first across both single-arm fine manipulation and dual-arm collaboration tasks in the benchmark, which measures zero-shot embodied AI performance in previously unseen environments.
According to Lumos, Prime R0 surpassed models including Nvidia’s 16-billion-parameter Cosmos model as well as entries from researchers at the Massachusetts Institute of Technology (MIT) and Princeton University, despite using less than one-sixth as many parameters.
The company says the result validates its strategy of developing embodied AI systems for industrial deployment rather than focusing on increasingly large foundation models.
Prime R0 has been integrated with the company’s Lumos Touch robotic arm and has been demonstrated in applications including floral arrangement, textile handling, precision parts storage and sorting materials in confined spaces.
Developed by the Allen Institute for AI (AI2), MolmoSpaces evaluates embodied AI models under standardized conditions using zero-shot inference in fixed environments. The benchmark measures how well models generalize across nearly 100 previously unseen environments, object categories and manipulation tasks.
Lumos says Prime R0 combines vision-language-action (VLA) decision-making with world-model-based prediction to allow robots to anticipate the physical consequences of actions before they are executed.
The model incorporates several proprietary technologies, including temporally adaptive action generation, unified geometric action representation, lightweight implicit physical prediction and mixture-of-experts (MoE) networks.
The company says Prime R0 was designed specifically for industrial environments, where factors such as reliability, inference speed, hardware cost and scalability are often more important than model size.
According to Lumos, the model can run locally on a consumer-grade Nvidia GeForce RTX 5060 8GB GPU, reducing hardware requirements compared with cloud-based systems. It also claims millisecond-level inference times and support for multiple industrial tasks, including picking, assembly, sorting and dual-arm collaboration.
Lumos says Prime R0 is built on its Lumos NexCore physical AI platform, which integrates industrial data, robotic hardware, foundation models and deployment tools into a closed-loop system designed for manufacturing environments.
Yu Chao, founder and CEO of Lumos Robotics, said: “Lumos NexCore is an operating system for industrial embodied intelligence, and Prime R0 is the first flagship model built on that foundation.
“We will continue expanding from manufacturing into logistics and additional sectors, with the goal of building the foundational platform for next-generation industrial robotics worldwide.”
The company says it views Prime R0’s performance on the MolmoSpaces benchmark as evidence that deployment-focused embodied AI can deliver practical results for industrial automation while supporting future expansion into logistics and other sectors.
