Physical AI Inference & Trajectory Upscaling for ros2_control

AI policies (VLAs, Diffusion Policy) output joint targets at 20–50 Hz, while robot hardware requires continuous commands at 500–2000 Hz. Naively forwarding sparse waypoints causes velocity discontinuities and acceleration spikes that trigger hardware faults or physical damage. Existing solutions are batch-oriented and cannot handle streaming, jittery AI output. This project builds a dedicated Inference-to-Control Bridge in ros2_controllers. It ingests streaming AI waypoints via topics and upscales them to hardware rate using C²-continuous quintic spline transitions, solving the transition problem. A lock-free look-ahead buffer absorbs inference jitter, a time-scaling mechanism handles buffer underrun gracefully, and a safety monitor enforces velocity, acceleration, and jerk limits via time dilation before every hardware dispatch. Deliverables: inference_bridge_controller plugin, quintic spline primitives in control_toolbox, full C++ test suite, and a Gazebo end-to-end demonstration on a simulated 6-DOF arm.