KPI：用 GPT-6 Astra 把自然语言指令转成人形机器人接触参数
这套人形机器人控制层在 Unitree G1 上运行，由 GPT-6 Astra 驱动的 agent 层把一句自然语言变成方向、跟踪、柔顺度、力范围等交互契约，无需任务专用代码。
NEW RESEARCH: A frozen prior can act as an efficient interaction middleware! Called KPI, short for A Promptable Kernel for Physical Interaction on Humanoids, it is a control layer that sits between whatever generates a robot's motion and an unmodified whole-body tracker, and decides how the robot should behave when it touches something. Its premise is that a trajectory alone says where to go but not how to push: at contact, the executing controller determines the outcome. KPI takes a four-field interaction contract (direction, tracking, compliance, force range, and allowed touch channels) and adapts stiffness, damping, reference and feedforward. The network works online at 100 Hz from tracking error and an estimated wrench, with the Cartesian impedance arm responding at 500 Hz and wrench estimation refreshing at 200 Hz. The wrench is estimated without adding any wrist force sensor. It runs on a Unitree G1 with SONIC as the untouched whole-body controller for legs and waist, and an agentic layer (Analyzer, Detector, Verifier, implemented with GPT-6 Astra) turns a single natural-language instruction into those contracts with no task-specific code. Demonstrated tasks include operating a winch to hoist a second robot, opening and walking through a door, lifting and placing boxes, writing on a board, teleoperated drawer opening and carrying while walking. What I find interesting, is that the unit of instruction is an interaction contract rather than a trajectory or an action. -> In practice, it is a frozen-backbone argument applied to control. The whole-body tracker (SONIC) is explicitly unmodified and the arm's impedance loop is untouched. KPI only modulates gains, reference and feedforward on top.