Alibaba unveils AI models for robots as China shifts focus to physical AI and autonomous systems

Alibaba Introduces Comprehensive AI Models for Robots

Alibaba has launched its first suite of AI models for robots, marking a strategic shift from conversational chatbots to physical AI systems capable of performing complex real-world tasks 1

1

. The Qwen-Robot Suite comprises three foundation models that form what the company describes as a "full stack for embodied intelligence" 2

2

. This move positions Alibaba as the only company in China spanning all five layers of the AI stack, from chips through cloud infrastructure, models, serving platforms, and applications built on top 1

1

.

The newly introduced models address the fundamental gap between vision and language understanding and physical control, which remains the central bottleneck for embodied AI 3

3

. At the center of this effort is RynnBrain, a system built to help machines understand space, objects, and motion—the perceptual groundwork a robot needs before it can act in the physical world 1

1

. Alongside the robotics suite, Alibaba announced Qwen3.7-Max, the latest in its proprietary large-language-model line, which can run autonomously for up to 35 hours without performance degrading 1

1

.

Three Core Models Define the Operating System for the Robot Economy

The Qwen-Robot Suite consists of three specialized models, each addressing distinct challenges in robotics. Qwen-RobotNav is a scalable vision-language navigation model that unifies five navigation tasks: instruction following, point-goal navigation, object search, target tracking, and autonomous driving 2

2

. Trained on 15.6 million samples, it achieves 76.5% success on VLN-CE RxR, a benchmark for vision-and-language navigation in real-world environments, and 90% tracking on EVT-Bench 2

2

.

Qwen-RobotManip is a generalizable vision-language-action model that tackles one of robotics' biggest challenges: different robots represent actions in fundamentally different ways 2

2

4

4

. To bridge these incompatible action spaces, Alibaba synthesized approximately 38,100 hours of training data from open-source robot datasets and human videos, without relying on proprietary data collection 2

2

. The model ranks first on RoboChallenge Table30-v1, outperforming previous approaches by 20% 2

2

.

Qwen-RobotWorld represents the most ambitious component: a video-based world model designed for embodied intelligence that treats natural language as a universal action interface 2

2

3

3

. The Embodied World Knowledge corpus spans 8.6 million video-text pairs—200 million frames—across manipulation, autonomous driving, indoor navigation, and human-to-robot transfer across 14 robot arms 2

2

. It ranks first on EWMBench and DreamGen Bench and scores perfectly on physics adherence, including Newton's laws, mass conservation, fluid dynamics, and gravity 2

2

.

Physical AI Push Reflects Strategic Industry Shift

The launch comes as Chinese firms, like their American counterparts, have concluded that the more lucrative business lies not in conversational models but in systems that can take actions—book, buy, operate, schedule—on a user's behalf 1

1

. Robotics is the most physical expression of that bet, extending AI agents from the screen into warehouses and homes 1

1

. The company's physical AI push aligns with a national strategy that treats both AI and robotics as priorities, pairing a domestic model stack with China's manufacturing base in a vertical play that software-only rivals find harder to match 1

1

.

These AI models for embodied intelligence have entered real-world pilot testing with select Alibaba Cloud enterprise customers within the robotics sector 4

4

. The launch came one week after Alibaba Group formed a new business unit known as Token Foundry, led by CEO Eddie Wu, combining the company's Tongyi Lab and Future Life business units to strengthen its AI efforts 3

3

. Wu has stated that Alibaba expects AI-related product revenue to become the primary driver of revenue growth for the cloud segment 4

4

.

What This Means for the Future of Autonomous Systems

While Western labs including Google DeepMind, Nvidia, Figure, and Physical Intelligence pursue similar goals, most focus on navigation or manipulation separately, not a unified, composable suite 2

2

. The distinction between language models and world foundation models is critical: while a language model predicts tokens, these models must understand physics, spatial relationships, and consequences of physical actions 2

2

. A language model tells you a glass breaks if dropped; Qwen-RobotWorld predicts how it breaks—shatter pattern, fluid dynamics, secondary collisions 2

2

.

The gap between controlled demonstrations and reliable real-world deployment remains enormous. The benchmarks these models excel on—RoboCasa365, LIBERO-Plus, RoboTwin-Clean2Rand—are simulation environments 2

2

. Real-world deployment introduces sensor noise, actuator drift, and the long tail of edge cases that have humbled every robotics effort in history 1

1

. Alibaba has not detailed pricing, availability, or which customers will receive the robot models first 1

1

. What the company has established is a position: a claim to span the whole stack at the moment the industry decides AI agents, not chatbots, are the prize 1

1

.