Alignment Unlocks Scale for Robotic Manipulation Foundation Models
Qwen Team
Welcome to the official repository of Qwen-RobotManip. Here, you can find official information about Qwen-RobotManip and post your questions (Issues).
Note: There is currently no plan to release the model weights for Qwen-RobotManip or Qwen-RobotNav. We will continue adding report resources that can be publicly released to this repository.
Qwen-Omni × Qwen-RobotManip — Qwen-Omni observes the scene, randomly proposes manipulation tasks via speech, and judges execution in real time. The two demos below are English-subtitled versions. Qwen-RobotManip completes tasks on the fly with no pre-defined task list, demonstrating open-ended instruction following and generalization.
clipboard.mp4 |
clipboard.mp4 |
Qwen-RobotManip is validated across real-robot platforms and tasks, demonstrating generalization to novel scenes, unseen language instructions, and cross-embodiment transfer.
clipboard.mp4 |
clipboard.mp4 |
clipboard.mp4 |
If the videos do not render in your browser, open the direct links: Omni Demo 1, Omni Demo 2, Real Robot 1, Real Robot 2, and Real Robot 3.
Qwen-RobotManip is a generalizable vision-language-action foundation model built upon Qwen-VL / Qwen3.5-4B. It couples a vision-language backbone with a flow-matching Diffusion Transformer action expert, enabling continuous action generation while preserving the perception and language grounding needed for robotic manipulation.
The central principle is alignment before scale. Robot manipulation data is naturally heterogeneous: robot embodiments, action spaces, camera systems, coordinate frames, collection pipelines, and task distributions vary widely. Qwen-RobotManip introduces a unified alignment framework across representation, motion, and behavior so that multi-source training becomes coherent instead of conflicting.
Using only open-source robotic manipulation datasets and egocentric human videos, without proprietary data collection, Qwen-RobotManip constructs a pretraining corpus of approximately 38,100 hours and demonstrates strong OOD generalization, instruction following, reactive recovery, and cross-embodiment transfer.
-
🔗 Three-Dimensional Alignment. Representation alignment, camera-frame motion alignment, and behavior alignment make heterogeneous manipulation data trainable under one model.
-
🤖 Unified Cross-Embodiment Action Space. An 80D canonical state-action representation with per-dimension binary masks accommodates single-arm, dual-arm, dexterous hand, and mobile manipulation settings.
-
📷 Camera-Frame End-Effector Motion. End-effector deltas are represented in the camera coordinate frame, with Camera Positional Encoding (CaPE) injecting camera geometry into the action expert.
-
🌍 Open-Source Data at Scale. The pretraining corpus contains about 38,100 hours of manipulation data, including 24,808 hours of human-to-robot synthetic demonstrations across 15 embodiments.
-
🏆 Strong OOD and Real-World Performance. Qwen-RobotManip leads OOD benchmarks such as LIBERO-Plus, RoboTwin-Clean2Rand, EBench, RoboCasa365, and RoboTwin-IF, and ranks #1 on the RoboChallenge Table30 v1 generalist track.
Qwen-RobotManip combines a Qwen3.5-4B vision-language backbone with a flow-matching DiT action expert. The VLM processes multi-view observations, language instructions, structured embodiment prompts, and execution context. The action expert predicts continuous action chunks through flow matching and uses a small number of Euler integration steps for low-latency control.
The model maps heterogeneous robot states and actions into a shared 80-dimensional canonical vector. Missing degrees of freedom are zero-padded and excluded from the loss with binary masks, allowing different embodiments to share one template without forcing nonexistent joints or grippers to receive supervision.
For motion alignment, Qwen-RobotManip predicts camera-frame delta end-effector poses, making visually similar motions numerically closer across embodiments and coordinate systems. Camera extrinsics and intrinsics are injected through CaPE, while end-effector type embeddings and auxiliary flags condition the denoising process.
For behavior alignment, Qwen-RobotManip uses structured embodiment prompts and in-context policy adaptation. Recent observation-state-action chunks from the same episode act as an implicit description of the current embodiment and behavior profile, allowing the model to adapt without parameter updates.
The Human-to-Robot synthesis pipeline converts egocentric hand demonstrations into robot demonstrations through action retargeting, hand removal and inpainting, simulated robot rendering, and depth-guided compositing. This pipeline turns about 1,933 hours of egocentric human video into about 24,808 hours of robot-compatible demonstrations across 15 robot embodiments.
Qwen-RobotManip matches or exceeds prior methods on standard in-distribution benchmarks, while the report emphasizes that IID results alone do not reliably measure foundation-model generalization.
| Model | LIBERO SR (%) | RoboTwin Easy SR (%) | RoboTwin Hard SR (%) |
|---|---|---|---|
| Qwen-RobotManip | 99.1 | 93.4 | 92.5 |
| Qwen-RobotManip-Context | 99.2 | 93.7 | 94.0 |
Qwen-RobotManip is evaluated across task and scene variation, instruction following, and cross-embodiment transfer.
| Benchmark | Metric | Qwen-RobotManip | Qwen-RobotManip-Context |
|---|---|---|---|
| LIBERO-Plus | Overall SR (%) | 89.0 | 91.4 |
| RoboTwin-Clean2Rand | Hard SR (%) | 62.6 | 69.4 |
| EBench | Overall SR (%) | 45.6 | 43.6 |
| RoboCasa365 | Total SR (%) | 35.9 | 33.8 |
| RoboTwin-IF | Average SR (%) | 72.2 | 72.0 |
RoboChallenge Table30 v1 further tests real-world OOD generalization under platform, task, and execution complexity shifts. The generalist track spans 30 tasks across 4 robot platforms, covering AgileX ALOHA, Franka, UR, and ARX. Qwen-RobotManip ranks #1 with 45% success rate and a 59.83 process score, outperforming the third-place system by 20%.
-
Bimanual coordination. On the 8 ALOHA tasks that require tight two-arm coordination, Qwen-RobotManip reaches 40% average success rate, compared with 21.2% for
$\pi_{0.5}$ . It is also the only model to succeed on the "pour fries into plate" task, which requires sequential stabilization, opening, lifting, and pouring. -
Cross-embodiment pick-and-place. Across 12 pick-and-place-centric tasks distributed over all four platforms, from single-object grasping to multi-step manipulation involving 4-5 objects, Qwen-RobotManip achieves 63.3% average success rate, surpassing the next-best baseline DM0 (48.3%) by 15.0 percentage points.
- Reactive recovery. In real-robot trials, the policy can retry after slips or failed grasps without an explicitly scripted recovery routine, suggesting that large-scale aligned pretraining helps the model close the loop between visual feedback and corrective action.
If you find our work helpful, feel free to give us a cite.
@misc{qwenrobotmanip2026,
title={Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models},
author={Qwen Team},
year={2026},
eprint={2606.17846},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2606.17846},
}