Learning Low-Frequency Motion Control for Robust, Dynamic Robotic Locomotion
Robotic locomotion is often approached with the goal of maximizing robustness and reactivity by increasing motion control frequency. Not in this video. Watch to see researchers challenge this intuitive notion by demonstrating robust and dynamic locomotion with a learned motion controller executing at as low as 8 Hz on a real ANYmal C quadruped.
More From SAE Media Group
Tech Briefs
Super-Fast Soft Robots Inspired by Cheetahs
Robotics & Automation INSIDER
A Springtail-Like Jumping Robot
Tech Briefs
Dynamic Fabric for Motion Generation
Tech Briefs
Torque Transmission Standout
Tech Briefs
Blur-Canceling Roll Camera for Robotics Testing
Motion Design INSIDER
Controller Helps Robots Avoid Collisions
Tech Briefs
5Ws of Film-Balloon Soft Robots
Motion Design
Rat-Like Nimble Robot Could Someday Perform Inspections
Motion Design
Precision Low-Speed Motor Controller
Tech Briefs
Neural Hand Prosthesis Control Test
Robotics & Automation INSIDER
Researchers Explore the Phenomenon of Holonomy
Motion Control and Automation Technology INSIDER
Mapping Human Movement to Improve Rehabilitative Robotics
Tech Briefs
A Picking Robot for a Greenhouse
Tech Briefs
Purdue Helps Robots Navigate Unstable Ground
Motion Design
Small Robot Has Outstanding Vertical Agility
Transcript
00:00:01 in this work we investigate low frequency motion control for legged robotic locomotion we challenge The High Frequency control design choice and show that robust and dynamic Locomotion Behavior can be achieved even at low control frequencies in this clip for example we demonstrate a blind Locomotion policy trained and simulation using a deep reinforcement learning
00:00:22 approach without any Dynamics randomization being executed on the real robots at 8 Hertz our research is motivated by the ability of animals to exhibit a child or commission skills despite significant sensory and actuation delays for a 40 kilogram animal for example the sensory motor latency corresponds to approximately 67.2 milliseconds which
00:00:46 considerably limits its motion control frequency we therefore trained and evaluated motion control policies to generate this adjoint position commands at frequencies ranging from 5 Hertz to 200 Hertz the motion control policies trained and deployed adverty here to more we observed aggressive actuation tracking resulting in significant motor
00:01:10 vibrations despite having been trained with an actuator Network which was used to model the real robot actuation Dynamics in simulation in contrast we observe stable and smooth Locomotion Behavior with motion control policies trained and deployed at 25 Hertz or less note that while we were able to transfer the five Hertz policy onto the real
00:01:41 robot we observed poor reactivity and therefore could not perform highly Dynamic Locomotion at five Hertz we posted that no frequency motion control is less sensitive to actuation Dynamics allowing policies during at low frequencies to operate as planners as opposed to controllers this further offers better robustness durations and system and environment Dynamics under
00:02:14 the assumption that the low level actuation controller performs stable tracking of desired joint States we refer to this as the mission Platinum hypothesis additionally we observe that at frequencies less than 25 Hertz and actually the network is not even required for a successful centurial transfer
00:02:35 we also demonstrate that I make Locomotion of uneven terrain is possible at a low motion control frequency of 8 Hertz the eight Hertz policy is also able to achieve a heading velocity of more than 1.5 meters per second