Seamless Reaction Strategy for Bipedal Locomotion Exploiting Real-Time Nonlinear MPC
KAIST DRCD Lab researchers present a real-time nonlinear model predictive control method for helping bipedal robots recover from disturbances using ankle, hip, and stepping adjustments.
In this talk
- 0:11
Combining walking responses
The method combines three ways to respond to pushes: ankle torque, hip motion that counters body rotation, and changes to the swing foot’s landing position or timing. The researchers formulate these responses together in a nonlinear model predictive control problem for bipedal walking.
- 1:35
Simulated push recovery
Simulations cover forward and sideways pushes during forward and sideways walking, including disturbances up to 23.5 newton-seconds. When footstep timing changes are disabled, the model withstands only 20.5 newton-seconds, a 13.6% reduction in recovery capacity.
- 3:10
Adapting to different robots and control weights
The method is also evaluated on point-foot and small-foot models that cannot use ankle responses. Changing the control weighting factors shifts the response among ankle action, hip motion, landing-position changes, and footstep-timing changes.
- 5:52
Real-robot tests
The controller runs on Gazelle, a 13-degree-of-freedom biped, with pushes measured by a force sensor. Tests report recovery from disturbances up to 46% above a baseline and a side push of 12.99 newton-seconds, 57.4% above baseline; solve times meet real-time requirements.
- 7:16
Walking over debris
In the final experiment, Gazelle walks with 0.15-meter strides over debris containing obstacles of different heights and objects that may move.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.