The model got faster, but can the robot move? An eight-step action-chunk adaptation experiment from OpenVLA-OFT to Qiyuan Q1

The story

An experiment adapted the 22-degree-of-freedom Qiyuan Q1 robot, based on OpenVLA-OFT, to use eight-step continuous action chunks with parallel decoding, continuous action representation, and L1 regression optimization. URDF simulation completed normalized mapping, joint-limit clipping, and rate constraints, while self-collision detection, dynamics validation, and real-robot closed-loop testing remain necessary.