Alibaba Launches Qwen-Robot Series: Embodied AI Models for Manipulation, Navigation, and World Modeling

Overview

On June 16, 2026, Alibaba announced the Qwen-Robot series — the first complete embodied intelligence model family from the Qwen large-model lineup. The release consists of three models:

• Qwen-RobotManip (VLA manipulation model) — giving robots "dexterous hands"
• Qwen-RobotNav (VLN mobility model) — giving robots "wayfinding feet"
• Qwen-RobotWorld (world model) — giving robots "a thinking brain"

The three models can be deployed independently or run collaboratively, offering a general-purpose foundation for different robot embodiments.

Source: Qwen official blog — https://qwen.ai/blog?id=qwen-robotmanip / qwen-robotworld / qwen-robotnav

Qwen-RobotManip: Making Robots Actually Manipulate Objects

Built on the Qwen-VL vision-language-action (VLA) model, it introduces a unified alignment framework covering representation, motion, and behavior. Using only open-source robot datasets and human demonstration videos, it was pretrained on roughly 38,100 hours of corpus spanning 15 robot embodiments.

Benchmark results:

• LIBERO-Plus: 91.4%
• RoboTwin-C2R Hard: 69.4%
• RoboCasa365 Composite-Unseen: SOTA

Key technical idea: instead of an end-to-end approach, it solves manipulation scalability through "alignment" — decomposing tasks into three mutually aligned subspaces, which is more interpretable and transferable than directly learning actions.

Qwen-RobotNav: Navigation Everywhere

Based on Qwen3-VL and trained on 15.6M samples, it unifies five domains: vision-language navigation, object-goal navigation, object tracking, autonomous driving, and embodied question answering. Tasks and observation parameters can be switched at inference time without architectural changes.

Benchmark results:

• VLN-CE RxR success rate: 76.5%
• HM3Dv2 object navigation: 75.6% (RGB-only)
• EVT-Bench tracking rate: 90.0%
• NAVSIM PDMS: SOTA

One model covering five task domains means robots deployed in factories, warehouses, homes, and autonomous vehicles can share the same navigation stack — substantially reducing engineering and maintenance costs.

Qwen-RobotWorld: Understanding How the World Works

It adopts a dual-stream Multimodal Diffusion Transformer (MMDiT) architecture with Qwen2.5-VL as the action encoder. Language serves as a unified action interface, standardizing over 500 action categories.

Key data points:

• Top results on 4 benchmarks
• Unifies 20+ robot embodiments
• 8.6M cross-scenario training pairs
• 1,300+ manipulation skills
• Joint training across manipulation, autonomous driving, and indoor navigation

The MMDiT architecture plus language interface lets robots describe target actions in natural language — valuable for human demonstration (e.g., "put the cup on the left side of the table") and cross-robot transfer (the same instruction given to a robotic arm, a humanoid, or an autonomous vehicle). It also supports Scene2Robot human-to-robot transfer.

Why It Matters

For robot developers: The three models are open-sourced together, letting teams fine-tune on Qwen-Robot instead of building a full embodied model stack from scratch — e.g., adding ~100 hours of proprietary data for domain adaptation on Qwen-RobotManip.

For China's embodied AI ecosystem: While NVIDIA GR00T, Google RT-X, and Physical Intelligence π0 have shipped overseas, models that truly land in Chinese industrial scenarios, support Chinese-language instruction interfaces, and target Chinese robot embodiments remain a gap. If the open-source license is friendly, Qwen-Robot could quickly become the de facto standard for domestic robots.

For the AI model landscape: Alibaba, ByteDance, Tencent, and Huawei have all entered embodied AI. Qwen-Robot expands "embodiment" from a single model to a complete system (manipulation + navigation + world model), setting a new product benchmark for the field.

For academia: All three models open training data and evaluation benchmarks (RoboTwin-C2R Hard, LIBERO-Plus, NAVSIM PDMS, etc.), enabling fair, like-for-like comparisons in follow-up papers.

Risks and Open Questions

1. Sim-to-real transfer: Strong benchmark scores don't guarantee real-world performance; robots at 90%+ success in labs often degrade sharply in unstructured factory or home settings. 2. Embodiment coverage: 15–20 embodiments sounds broad, but industrial robot categories vary widely (collaborative arms, SCARA, Delta, AGV, humanoids); real deployment still requires substantial adaptation. 3. Open-source license and commercialization: License terms, commercial use, and redistribution restrictions will determine whether Qwen-Robot becomes a true standard foundation. 4. Multi-model coordination: How Manip, Nav, and World synchronize state, pass context, and resolve decision conflicts in real deployments is not fully disclosed. 5. Gap with international players: NVIDIA GR00T N1 and Physical Intelligence π0 have trained on larger-scale data; Qwen-Robot needs time to prove competitiveness in data scale and diversity.

Bottom Line

Alibaba's "hands + feet + brain" trio pushes embodied AI from single-point breakthroughs toward a full-stack system. Whether Qwen-Robot becomes the de facto standard for domestic robots, the second half of 2026 will be the key observation window.