A quadruped robot walks just as well when its brain thinks eight times a second instead of two hundred.
Researchers trained reinforcement learning controllers for the ANYmal C quadruped robot at frequencies ranging from 5 Hz to 200 Hz. The 8 Hz version matched its higher-frequency counterparts, hitting a heading speed of 1.5 m/s, crossing uneven terrain, and shrugging off unexpected shoves. The team found low-frequency policies tolerate actuation delays and shifting system dynamics so well that sim-to-real transfer worked without the usual dynamics randomization or detailed actuator modeling. Training and deployment code, plus an extended analysis, are public.
Robotics has long assumed that faster control loops mean better balance, pushing teams toward pricier high-frequency hardware and heavier onboard compute. This result argues the opposite for legged robots: simpler, slower-cycling controllers can be just as robust, which matters for cost, power draw, and compute budgets on robots meant to actually ship.
It is a reminder that in robot locomotion, more compute per second is not automatically a better product - just a bigger one.