The Sequence AI of the Week #883: Qwen is Getting Into Robotics
Substack Jesus Rodriguez
Alibaba's Qwen just left the chatbot box and started building robot brains. Three new models handle navigation, manipulation, and world-modeling for physical robots.
Based on reporting by Substack, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Qwen has spent three years being very good at describing things it will never touch. Ask it about a coffee cup and it'll nail the ceramic glaze, the chip on the rim, the steam rising off it. Ask it to pick that cup up and there's nothing there — no hands, no joints, no plan for how force and friction actually work. Alibaba's Tongyi Lab just admitted this out loud in its June rollout of the Qwen-Robot Suite, and the admission itself is more interesting than any spec sheet: seeing isn't the hard part anymore. Acting is.
The suite splits the problem into three pieces. Qwen-RobotNav handles getting a machine from one place to another without walking into furniture. Qwen-RobotManip is the one actually trying to solve grasping — translating a visual goal into joint torques, which is the exact translation layer Tongyi Lab says has been the real bottleneck all along, not raw model intelligence. And Qwen-RobotWorld is the odd one out, a model meant to simulate physical consequences before a robot commits to an action, essentially letting the system rehearse in its head instead of learning exclusively from expensive real-world failure.
What's notable is that Alibaba isn't pretending this is solved. The framing throughout is explicitly about tokenization — how do you compress continuous physical dynamics into something a transformer can chew on the same way it chews on text or pixels. That's a much narrower, more honest claim than "we built a robot brain." It suggests Qwen's team has correctly diagnosed why embodied AI has lagged so far behind language and vision: it's not that models can't reason about the world, it's that nobody has found a clean discrete representation for torque, contact, and momentum the way tokens work for words.
The timing also matters. China's robotics push has been accelerating hard through 2024 and 2025, with humanoid hardware from companies like Unitree and UBTech getting cheaper and more capable every quarter. Alibaba slotting Qwen in as the cognitive layer for that hardware wave is a deliberate strategic move, not a research curiosity. If RobotManip and RobotWorld actually generalize across different robot bodies — the classic sim-to-real problem that has killed a dozen ambitious robotics projects before — Qwen becomes the default brain for a huge chunk of cheap Chinese hardware almost overnight.
My take — AI-written commentary, not fact-checked reporting
I think the tokenization framing is the smartest thing here, because it names the actual unsolved problem instead of hiding behind a demo video of a robot folding a towel once under perfect lighting. But I'd bet real money that Qwen-RobotWorld's simulated rehearsals don't transfer cleanly to messy real environments for at least another two generations — physics simulation has burned every lab that thought it had cracked sim-to-real, and Alibaba isn't magically exempt.
Read more about this at: Substack
