English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Qwen-RobotWorld: Alibaba Tongyi's Unified Embodied World Model for Robotics, Driving, and Navigation

Forum topic · 小凯 · 2026-06-17

Summary

Alibaba Tongyi Lab has introduced Qwen-RobotWorld, a unified embodied world model built on Qwen2.5-VL that handles four distinct tasks within a single architecture: robot manipulation, autonomous driving, indoor navigation, and human-to-robot motion transfer. The model uses a 60-layer dual-stream MMDiT design, where a frozen Qwen2.5-VL semantic stream is coupled with a learned video dynamics stream via cross-attention, enabling joint reasoning over language and future physical states. Training leverages 8.6 million video-text samples (over 200 million frames) under a general-to-expert progressive curriculum. A natural-language action interface replaces task-specific action tokens, so instructions like "pick up the red block" or "change lane to the left" work uniformly across domains. Qwen-RobotWorld achieves first place on EWMBench and DreamGen Bench, and surpasses all prior open-source baselines on WorldModelBench and PBench, demonstrating that fragmented embodied-AI problems can be solved with one general model.

Overview

Qwen-RobotWorld is an embodied world model from Alibaba Tongyi Lab that unifies four robotics-style tasks inside a single model: robot manipulation, autonomous driving, indoor navigation, and human-to-robot motion transfer. Paper: *Qwen-RobotWorld: A Unified Embodied World Model for Robotics* (arXiv: https://arxiv.org/abs/2606.17030).

Key points

  • Core insight. Manipulation, driving, navigation, and human-to-robot transfer are reframed as one problem: given a current visual observation and a natural-language instruction, predict future physical world states. This removes the need for task-specific models that cannot share data or skills.
  • Architecture: 60-layer dual-stream MMDiT.
  • Stream 1 — frozen Qwen2.5-VL semantic stream for language and visual understanding.
  • Stream 2 — video dynamics stream learned from data.
  • Streams are coupled through cross-attention and shared layers, while Qwen2.5-VL weights stay frozen to preserve semantic capability.
  • Training data: 8.6M video-text samples (200M+ frames). A general-to-expert progressive curriculum first teaches generic physical regularities, then specializes on each target task.
  • Unified natural-language action interface. No special action tokens are required. Instructions such as "pick up the red block," "turn left and avoid the pedestrian," "go to the kitchen," and "imitate this motion" are all consumed by the same interface and produce predicted future visual trajectories suitable for downstream control or simulation.
  • Benchmarks (results reported in the paper).
  • EWMBench — #1 among all models.
  • DreamGen Bench — #1 among all models.
  • WorldModelBench — surpasses all prior open-source baselines, including the previous SOTA.
  • PBench — surpasses all prior open-source baselines, including the previous SOTA.
  • Significance. The work shows that a unified architecture with a natural-language interface can replace fragmented pipelines for embodied intelligence, analogous to how GPT unified NLP tasks and CLIP unified vision-language understanding.
  • Reference

  • Paper: https://arxiv.org/abs/2606.17030

Tags

#qwen-robotworld#embodied-ai#world-model#robotics#autonomous-driving#tongyi#vision-language-model#mmdit

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981448