- The 2.8B-parameter model tops global benchmark, beating Nvidia and leading US academic research teams
- Designed for production environments, the model prioritizes productivity, speed and reliability over scale
Lumos Robotics (鹿明机器人), a Suzhou-based embodied AI company, has claimed the top spot on the latest MolmoSpaces leaderboard, a globally recognized benchmark for zero-shot embodied AI, with its industrial embodied foundation model Prime R0.
Prime R0 achieved the highest overall success rate across both single-arm fine manipulation and dual-arm collaboration tasks, outperforming leading competitors including Nvidia’s Cosmos (16B), MIT, and Princeton University.
Despite having only 2.8 billion parameters—less than one-sixth the size of Nvidia Cosmos—Prime R0 delivered the best overall performance. The achievement also marks one of the strongest international benchmark performances by a Chinese-developed embodied AI model and validates Lumos’ industrial-first approach to embodied intelligence.
From lab demos to industrial deployment
While much of the embodied AI industry remains focused on general-purpose demonstrations, Lumos believes the next stage of industry growth depends on deploying robots into production-ready environments where they can solve practical problems and generate measurable productivity gains.
To support this vision, Lumos has built a closed-loop ecosystem spanning industrial data, foundation models, robot hardware, and deployment scenarios. Developed within this framework, Prime R0 prioritizes generalization, responsiveness, and reliability in complex industrial settings rather than pursuing ever-larger parameter counts.
Integrated with the Lumos Touch robotic arm, Prime R0 has already demonstrated strong performance across demanding real-world applications, including sequential floral arrangement, textile manipulation, precision parts storage, and sorting disorganized materials in confined spaces.
Outperforming larger models
Developed by the Allen Institute for AI (AI2), MolmoSpaces is regarded as one of the most rigorous zero-shot embodied AI benchmarks.
Models are evaluated under strict conditions, including zero-shot inference, fixed environments, and standardized success-rate metrics.
The benchmark assesses a model’s ability to generalize across nearly 100 unseen environments, diverse object categories, and varied manipulation tasks. In this evaluation, Prime R0 surpassed significantly larger models despite operating with only 2.8B parameters, demonstrating the effectiveness of Lumos’ industrial embodied AI architecture in highly generalized scenarios.
Architecture-level innovation
Prime R0 combines Vision-Language-Action (VLA) decision-making with world-model-based prediction, enabling the system to anticipate the physical consequences of actions before execution. This creates a unified loop integrating perception, reasoning, and action.
The model is powered by four proprietary technologies:
- temporally adaptive action generation
- unified geometric action representation
- lightweight implicit physical prediction
- mixture-of-experts (MoE) networks
Together, these capabilities move beyond traditional perception-driven robotics and form the foundation of Prime R0’s industry-leading performance.
Many embodied AI systems excel in demonstrations but struggle to meet factory requirements around cost, reliability, cycle time, and scalability. Prime R0 was designed from the outset to address these industrial constraints.
Key advantages include:
- Over 80% lower hardware costs: Prime R0 can run locally on a consumer-grade NVIDIA GeForce RTX 5060 8GB GPU, eliminating dependence on expensive cloud-based compute infrastructure.
- Millisecond-level inference: Unlike many world-action models that require 5–6 seconds per inference step, Prime R0 delivers near real-time responses suitable for continuous production environments.
- Multi-task and multi-hardware adaptability: A single model supports picking, assembly, sorting, and dual-arm collaboration, while natively supporting Franka robotic arms without retraining.
- Higher operational efficiency: Leveraging SE(3) geometric priors and hybrid real-virtual training data, Prime R0 maintains strong performance under industrial challenges such as occlusion, stacking, and positional deviations, reducing manual intervention and improving overall equipment effectiveness (OEE).
Lumos views Prime R0 not as an isolated model breakthrough, but as the result of a broader full-stack strategy spanning robotic hardware, industrial data systems, physical AI engines, foundation models, and deployment toolchains.
At the center of this strategy is Lumos NexCore, a physical AI platform designed to continuously connect data, models, and real-world deployments through a closed-loop architecture. Unlike approaches focused primarily on academic research or cloud-scale computation, Lumos NexCore is built specifically for industrial deployment and guided by four principles:
- industrial value first
- full-domain system foundation
- physical intelligence at the core
- full-stack in-house development
“Lumos NexCore is an operating system for industrial embodied intelligence, and Prime R0 is the first flagship model built on that foundation,” said Yu Chao, Founder and CEO of Lumos Robotics. “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.”
Prime R0’s success on MolmoSpaces represents more than a benchmark victory. It demonstrates how deployment-driven innovation can make embodied intelligence practical, scalable, and ready for real-world industrial environments.
For Lumos, it is the beginning of a broader mission to build the foundational infrastructure for industrial embodied intelligence globally.
*This is a press release provided by Lumos Robotics. Please visit https://www.lumosbot.tech/ for further information.
