Qwen Drops RobotSuite: 3 Open Models for Robots

Qwen team at Alibaba has published Qwen-RobotSuite, consisting of three open embodied-AI models targeting different levels of the robotics stack: RobotManip for manipulation, RobotWorld for video world modeling, and RobotNav for navigation. Of the three, RobotManip and RobotNav are available on public Github repos, whereas RobotWorld is published as a research paper currently. This is intentional, because robotics is full of isolated models for specific tasks and robots, and Qwen is trying to offer a family of solutions covering the key roles.

RobotManip is a vision-language-action model based on Qwen3.5-4B and capable of generating continuous actions for robots. It’s pre-trained on approximately 38,100 hours of data from open sources and human videos only (no proprietary data), achieving SOTA on RoboChallenge Table30-v1 with a 20% relative improvement compared to the state of the art. Its secret sauce lies in an 80-dimensional canonical action vector with dynamic masking, allowing models from different robots to learn from each other without interference. In terms of cross-embodiment transfer, it achieves 23.9% success rate, which is around 3.2x higher than the 7.5% of the π0.5 baseline.

1Three jobs, one family

RobotWorld is a beast, consisting of a 20-billion parameter, 60-layer multimodal diffusion transformer for forecasting future frames given a language description, trained on 8.6 million video-language pairs across over 200 million observations. It leads on both EWMBench and DreamGen Bench overall, and on WorldModelBench among open models. RobotNav is available in 2B/4B/8B variants, generating 8-waypoints trajectories, and delivering decent performance on benchmark tasks such as 76.5% success rate on VLN-CE RxR and 75.6% on HM3D ObjectNav, along with an “agentic” mode reducing steps by 77% on one embodied-question benchmark.

2Why it matters to you

Most of us don’t care much about individual scores on these leaderboards, but we know that the open-source movement that transformed chatbots is coming to robots. Making source code available instead of just showing off a fancy demo lowers the threshold for universities, startups and DIY enthusiasts who cannot afford licensing a proprietary robotics framework. Two years ago, we saw the chatbot industry picking up thanks to public availability of good weights and code, enabling thousands of people to make improvements simultaneously instead of relying on a few vendors.

What I’m really excited about is that this model was trained on 38,100 hours of data without using any proprietary data. If other robot models can also be trained on public data and human videos, then the playing field becomes less about who has more robot logs, but who can better design the training process, which usually leads to faster progress across the board.