Alibaba has entered the race to build “physical AI” with a new model designed to power robots in the real world.
The Chinese tech giant on Tuesday unveiled RynnBrain, an artificial intelligence system built to help machines understand space, objects, and motion.
The launch signals Alibaba’s growing push into robotics, an area already attracting heavy investment from U.S. players like NVIDIA and Google.
A short demonstration video released by Alibaba’s DAMO Academy shows a robot recognizing fruit and placing it into a basket. The task appears simple.
The intelligence behind it is not. The robot must identify objects, track their positions, and plan precise movements in real time.
Alibaba positions RynnBrain as a foundational “embodied AI” model. That category includes robots, autonomous vehicles, and other machines that interact directly with the physical environment.
China has made physical AI a national priority as competition with the United States intensifies.
Built for real-world memory
RynnBrain focuses on solving a major weakness in existing robotics models: poor memory of space and time.
Traditional embodied AI systems often forget object locations or misinterpret scenes. Alibaba says RynnBrain addresses both problems through spatiotemporal memory.
That capability lets robots recall where objects appeared earlier and predict how they will move next.
The system also supports global retrospection. A robot can review past actions before choosing its next step. That approach helps reduce errors during complex tasks.
Physical-space reasoning adds another layer. RynnBrain combines text-based logic with spatial cues. This hybrid method allows robots to reason in ways that better reflect real-world environments.
Alibaba trained the model on its Qwen3-VL visual-language system. DAMO Academy optimized it using a custom architecture called RynnScale.
The company says the setup doubled training speed without increasing computing resources.
Smaller model, faster robots
The headline release includes what Alibaba calls the industry’s first 30-billion-parameter mixture-of-experts embodied AI model.
Despite its size, the system activates only 3 billion parameters during inference. Alibaba claims this efficiency allows the model to outperform much larger 72-billion-parameter systems.
Lower inference demands translate into smoother robot motion and faster decision-making.
Those gains matter for real-world deployment, where power and latency limits often restrict performance.
Performance tests suggest strong results.
Alibaba reports that RynnBrain set new records across 16 open-source embodied AI benchmarks. These tests measured environmental perception, spatial reasoning, and task execution.
The company says RynnBrain surpassed competing systems such as Google‘s Gemini Robotics ER 1.5 and NVIDIA’s Cosmos Reason 2.
Alongside RynnBrain, DAMO Academy released seven fully open-source models. These include base models and fine-tuned versions designed for commercial use.
Alibaba says the releases aim to lower research barriers for robotics developers.
Open access could speed adoption across manufacturing, logistics, and service robotics.
DAMO Academy also introduced a new evaluation framework called RynnBrain-Bench.
The benchmark targets a long-standing gap in embodied AI testing.
It focuses on fine-grained spatiotemporal tasks rather than static image recognition.
The move places Alibaba deeper into a global competition that includes U.S. tech leaders and startups.
As robots move beyond labs and simulations, embodied AI models like RynnBrain may define how machines operate in the physical world.
For now, Alibaba wants a seat at that table.
date: 2026-02-11 09:21:00