aiqbench analysis
This research introduces AVP, an architecture that decouples cognitive perception from motor control in Vision-Language-Action models to improve robotic manipulation efficiency. By using visual-primitive tokens conditioned by a VLM rather than forcing the action expert to relearn perceptual tasks, the method achieves significant gains in data efficiency and generalization on real robots. The approach represents a meaningful architectural shift for scaling embodied AI systems beyond single-task demonstrations.
Action with Visual Primitives
arXiv:2605.22183v3 Announce Type: replace-cross Abstract: Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalist robotic manipulation. A common design in current architectures maps language instructions and visual observations to actions in a single forward pass. While conceptually simple, this formulation entangles instruction comprehension, spatial scene understanding, and motor control within a single learning objective. As a result, the action expert must implicitly relearn cognitive and perceptual capabilities already present in the pretrained VLM, which can limit both learning efficiency and generalization. We introduce AVP (Action with Visual Primitives), an end-to-end architecture that implements this visual-primitive-centric interface: the VLM infers the next-stage target and emits visual-primitive tokens that condition a flow-matching action expert, with supervision derived from end-effector kinematics. Real-robot experiments on general pick-and-place tasks show that AVP improves the success rate by 37.04% over pi_0.5 and outperforms other recent methods, with consistent gains in data efficiency, spatial-compositional generalization, and object-level transfer.
Abstract:Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalist robotic manipulation. A common design in current architectures maps language instructions and visual observations to actions in a single forward pass. While conceptually simple, this formulation entangles instruction comprehension, spatial scene understanding, and motor control within a single learning objective. As a result, the action expert must implicitly relearn cognitive and perceptual capabilities already present in the pretrained VLM, which can limit both learning efficiency and generalization. We introduce AVP (Action with Visual Primitives), an end-to-end architecture that implements this visual-primitive-centric interface: the VLM infers the next-stage target and emits visual-primitive tokens that condition a flow-matching action expert, with supervision derived from end-effector kinematics. Real-robot experiments on general pick-and-place tasks show that AVP improves the success rate by 37.04% over pi_0.5 and outperforms other recent methods, with consistent gains in data efficiency, spatial-compositional generalization, and object-level transfer.
9 pages, 6 figures. Project page: this https URL
Robotics (cs.RO); Artificial Intelligence (cs.AI)
arXiv:2605.22183 [cs.RO]
(or arXiv:2605.22183v3 [cs.RO] for this version)
Full text under the publisher's open licence · 170 words
Related coverage
Action with Visual Primitives
arXiv · Sep 22, 2026
WOVEN: Weaving Visual World Modeling into Multimodal LLMs
arXiv · Oct 9, 2026
Latent Actions from Factorized Transition Effects under Agent Ambiguity
arXiv · Aug 3, 2026
[AINews] Opus 5.5 is good at explainer videos
Latent Space · Sep 29, 2026