PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control
PredActor runs a diffusion-based humanoid controller onboard, using only proprioception while still allowing text, joystick, and test-time steering.
The paper’s policy jointly predicts executable actions and an internal future-state trajectory, but executes only the actions. It avoids a separate motion-reference tracker and does not use externally estimated full-body state inputs. In simulation, it reached all 15 destination targets and improved text retrieval score over conditional action diffusion, with similar disturbance survival. The authors report runtime optimizations bringing the control callback under a 20 ms period on a Jetson Orin NX, and deployment on a Unitree G1. ArXiv · AI/CL/LG's note
The paper’s policy jointly predicts executable actions and an internal future-state trajectory, but executes only the actions. It avoids a separate motion-reference tracker and does not use externally estimated full-body state inputs. In simulation, it reached all 15 destination targets and improved text retrieval score over conditional action diffusion, with similar disturbance survival. The authors report runtime optimizations bringing the control callback under a 20 ms period on a Jetson Orin NX, and deployment on a Unitree G1. ArXiv · AI/CL/LG's note
score 5