Alibaba RynnBrain: New AI Model Powers Robots with Enhanced Spatial Awareness

Alibaba is making a significant push into the rapidly evolving field of “physical AI” with the launch of RynnBrain, a new artificial intelligence model designed to give robots a more robust understanding of the physical world. The unveiling, Tuesday, signals the Chinese tech giant’s ambition to compete with established players like Google and NVIDIA in the development of AI systems capable of powering real-world robotics.

At its core, RynnBrain aims to address a critical limitation in current robotics AI: the difficulty in maintaining a consistent and accurate perception of space, objects, and motion over time. Traditional embodied AI systems often struggle with “forgetting” object locations or misinterpreting dynamic scenes. Alibaba’s solution centers around what it calls “spatiotemporal memory,” enabling robots to recall past observations and predict future movements.

A demonstration released by Alibaba’s DAMO Academy illustrates this capability with a seemingly simple task: a robot identifying fruit and placing it into a basket. While the action itself is straightforward, the underlying intelligence is complex. The robot must not only recognize the fruit but also track its position, plan a precise trajectory, and execute the movement in real-time – all while accounting for potential changes in the environment.

RynnBrain is positioned as a foundational “embodied AI” model, a category encompassing robots, autonomous vehicles, and other machines that directly interact with the physical environment. This focus aligns with a national priority in China, which is increasingly viewing physical AI as a key area for technological advancement and competition with the United States.

Built for Real-World Persistence

The model’s spatiotemporal memory isn’t simply about remembering *where* things are, but also *when* they were there and how they’re likely to move. This allows for more reliable performance in dynamic environments. Alibaba also highlights the system’s “global retrospection” capability, allowing robots to review past actions before making new decisions, a feature designed to minimize errors during complex tasks. This is coupled with “physical-space reasoning,” which combines text-based logic with spatial cues, allowing robots to reason about their surroundings in a way that more closely mirrors human understanding.

RynnBrain was trained on Alibaba’s Qwen3-VL visual-language system and optimized using a custom architecture called RynnScale. According to Alibaba, this optimization doubled training speed without requiring additional computing resources, a significant efficiency gain.

Smaller Model, Faster Robots

A key aspect of RynnBrain’s design is its efficiency. Alibaba describes it as the industry’s first 30-billion-parameter mixture-of-experts embodied AI model. However, crucially, only 3 billion parameters are active during inference – the process of the model making predictions. This selective activation allows RynnBrain to outperform larger, 72-billion-parameter systems, according to Alibaba’s internal testing.

This reduced computational demand translates directly into smoother robot motion and faster decision-making. In real-world applications, where power consumption and latency are often critical constraints, this efficiency is a major advantage. Lower inference demands mean robots can react more quickly and reliably to changing conditions.

Alibaba reports that RynnBrain achieved top scores across 16 open-source embodied AI benchmarks, measuring environmental perception, spatial reasoning, and task execution. The company claims it surpassed competing systems, including Google’s Gemini Robotics ER 1.5 and NVIDIA’s Cosmos Reason 2. While independent verification of these claims is pending, the results suggest a significant step forward in embodied AI performance.

Alongside RynnBrain, DAMO Academy has released seven fully open-source models, including base models and fine-tuned versions intended for commercial use. This move is intended to lower the barrier to entry for robotics developers and accelerate the adoption of embodied AI across various industries, including manufacturing, logistics, and service robotics.

To further support the development community, DAMO Academy also introduced RynnBrain-Bench, a new evaluation framework designed to address a gap in existing embodied AI testing methodologies. Current benchmarks often focus on static image recognition, while RynnBrain-Bench emphasizes fine-grained spatiotemporal tasks, providing a more realistic assessment of a robot’s ability to operate in a dynamic environment.

The launch of RynnBrain places Alibaba squarely in a global competition with U.S. Tech leaders and emerging startups. As robots transition from controlled laboratory settings to real-world deployments, embodied AI models like RynnBrain will play a crucial role in defining how these machines operate and interact with their surroundings. Alibaba’s ambition is clear: to secure a prominent position in this evolving landscape.