Tencent Hy-Embodied-0.5-VLA-RoboTwin

Hy-Embodied-0.5-VLA-RoboTwin is what Tencent's general-purpose action model becomes after it is trained for one specific set of tasks. The starting point is Hy-Embodied-0.5-VLA-UMI; the finish is a checkpoint fine-tuned on all fifty two-armed manipulation tasks of the RoboTwin 2.0 benchmark, 550 recorded attempts per task, on 32 GPUs. The reported result is 90.9% success on the clean version of the benchmark and 90.1% when the scene is randomised — object positions, lighting and clutter shuffled. Tencent presents this as the best published figure for a vision-language-action model. The gap of less than one point between the tidy and the disturbed setting is the more informative number: it suggests the policy is not memorising exact positions. Two things separate it from its own starting point. It watches six frames instead of one — the current image plus five from the recent past — so it can tell whether an object is already moving. And its planning horizon is shorter in steps but longer in time: twenty actions at a third of the rate, covering more seconds of the future than the fifty fast steps of the pre-trained version. The honest caveat belongs in the same breath as the score. RoboTwin 2.0 is a simulated benchmark, and this checkpoint was tuned for its fifty tasks specifically. A high number here is evidence that the underlying method works, not that this file will make a robot useful in a real kitchen. For that, Tencent points users back to the pre-trained model and their own data. The licence is Apache 2.0 — worth noting, because the vision-language backbone this family is built on carries a Tencent licence that expressly does not apply in the European Union, the United Kingdom or South Korea. Technical report: arXiv 2606.14409.

▸Videos

No videos yet.

▸Reviews