Ten days ago, NVIDIA released Cosmos 3 with the opposite bet: collapse VLM reasoning, video simulation, and action generation into one 64B-parameter model with joint attention between the two towers. Alibaba has now come back with a different answer. The Qwen-Robot Suite, released June 16, breaks the problem into three distinct models — and that design decision reflects a real disagreement about where the bottlenecks in embodied AI actually are.
The three models are Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld. Each targets a different part of the physical action space. Nav handles mobile robots: instruction following, goal-directed navigation, target tracking, and autonomous vehicle tasks — a meaningful set of subtasks that share the property that the robot’s body moves through space rather than reaching into it. Manip handles arm-type systems: grasping, placing, and object manipulation where the end-effector interacts with the world directly. World is a predictive model: given an observation and a natural-language-described action, it generates a prediction of the resulting physical state, applicable across navigation, driving, and manipulation scenarios.
The cross-embodiment problem is the part worth focusing on. Training a robotics model has historically been painful because robot hardware is fragmented — different joint configurations, different sensor layouts, different degrees of freedom, different control frequencies. A model trained on one robot’s data often transfers poorly to another platform even when the task is similar. The Manip model’s approach to this is to project heterogeneous robot data into a canonical state-action space, where the representation is decoupled from the specifics of any particular hardware. The 38,100 hours of open-source manipulation data it trains on came from diverse platforms; the canonical projection is what allows all that data to cohere.
That’s a meaningful technical contribution independent of how well the individual models perform. Embodied AI has a data problem — quality labeled data from real physical interactions is expensive, time-consuming, and mostly siloed within individual research groups. Anything that increases the leverage of existing public data is structurally useful.
The World model is doing something adjacent to what Cosmos 3’s Generator tower does, but through a different mechanism. Rather than joint attention between reasoning and generation paths in a single model, Qwen-RobotWorld operates as a standalone system that learns to predict future physical states as a latent representation. The claim is that this gives Nav and Manip models a way to reason about consequences without needing end-to-end fusion — they can query the world model separately during planning rather than requiring physics understanding to be baked into every forward pass of the action-generation model.
Whether that decomposed approach works better in practice than Cosmos 3’s joint training is genuinely unclear, and probably task-dependent. For mobile navigation over novel terrain, having a separate world model available for look-ahead planning seems plausible. For fine-grained manipulation where the physical dynamics of a specific grasp are what matters, it’s less obvious that a separate world model gives you what the policy network needs at inference time.
Alibaba hasn’t published detailed benchmarks alongside the launch, which makes it hard to compare directly against Cosmos 3 or other embodied AI work. What’s available are the functional descriptions and the open-source data claim for Manip. The Qwen-VLA page connects this to the broader Qwen vision-language work that underpins all three models.
The architectural bet — specialized models with a shared foundation and a canonical representation layer — is a reasonable one. The counterargument is that you end up with three models to maintain, three inference pipelines to run, and integration overhead whenever a task spans the boundaries between them (a mobile robot that also manipulates objects, say). The counterargument to the counterargument is that task-specific models can be optimized, quantized, and deployed differently depending on the hardware constraints of the actual robot, whereas a single 64B model running on every platform is harder to work with at the edge.
For practitioners building embodied AI systems, the release is worth watching if only because the Manip model’s cross-embodiment training approach could make the public manipulation data substantially more usable. Whether the specific models clear the bar for real-world deployment is something that will take independent evaluation to determine.