[IEEE 2008 IEEE International Conference on Robotics and Automation (ICRA) - Pasadena, CA, USA...
Transcript of [IEEE 2008 IEEE International Conference on Robotics and Automation (ICRA) - Pasadena, CA, USA...
![Page 1: [IEEE 2008 IEEE International Conference on Robotics and Automation (ICRA) - Pasadena, CA, USA (2008.05.19-2008.05.23)] 2008 IEEE International Conference on Robotics and Automation](https://reader036.fdocuments.in/reader036/viewer/2022080420/5750a3d71a28abcf0ca5c5a2/html5/thumbnails/1.jpg)
Learning Capture Points for Bipedal Push Recovery
John R. Rebula, Fabian Canas, Jerry E Pratt, Ambarish Goswami
Researchers at IHMC and Honda Research Institute are de-veloping techniques for Learning Capture Points for BipedalPush Recovery. A Capture Point is a point on the groundwhere a biped can step to in order to stop. Humans arevery adept at stepping to Capture Points, while most bipedalrobots cannot recover from significant pushes.
To calculate approximate Capture Point locations, we usethe Linear Inverted Pendulum Model introduced by Kajitaand Tani. For a point mass biped walking at a constant height,this model exactly predicts the Capture Point. However, fora distributed mass biped, it is only an approximation. Inorder to better predict Capture Points, we learn a correctionfunction to the Linear Inverted Pendulum Model. A LinearInverted Pendulum maintains a conserved quantity called theorbital energy throughout the stance phase, which traces outtrajectories on the phase plot. The equation for orbital energycan be used to calculate a Capture Point by calculatingthe current energy, predicting the velocity at the supportfoot transfer, and then setting the next swing orbital energyto zero and solving for the body position. We used twolearning methods, one online and one offline, to improveCapture Point prediction. In the offline learning method, therobot is pushed multiple times with a given force magnitudeand direction. During each learning trial, the robot startsbalanced upright and stationary. Pushes, represented in thevideo as projectiles, are applied as a constant force for 0.05seconds. We vary the magnitude of the force from 150 to550 Newtons, and its direction in the horizontal plane. Whenthe Center of Mass reaches a set horizontal distance fromthe support foot, the robot predicts a Capture Point andstarts to step to that location. For each applied force, severalsimulation runs are performed, each one resulting in the robottaking a step within a grid centered around the Capture Pointpredicted by the Linear Inverted Pendulum Model.
We find multiple capture points because the actuated feetallow for a margin of error. These points define a captureregion. For failed points we calculate the center of massposition and velocity vectors at the top of swing. Theseare then used to calculate the residual Orbital Energy. Fortrials that result in successfully stopping, the residual OrbitalEnergy is zero.
This process is repeated for five different forces at 10different angles around the circle. For each of these pusheswe calculate the centroid of the capture region and the
J. Rebula, J. Pratt, and F. Canas are with the Florida Institutefor Human and Machine Cognition, Pensacola, Florida 32502, [email protected]
A. Goswami is with the Honda Research Institute Mountain View, CA94041, USA [email protected]
offset vector from the Capture Point predicted by the LinearInverted Pendulum model.
We then fit a curve to the angle and magnitude of theoffset vector as a function of the angle of the center ofmass velocity. Using this corrected capture point, we findsignificant reduction in residual energy error when comparedto using only the Linear Inverted Pendulum Model.
In the online learning technique, we use a Radial BasisFunction to represent the learned offsets from the CapturePoint predicted by the Linear Inverted Pendulum Model.After each push, the robot predicts where to step to, stepsthere, and determines the residual orbital energy error. Weuse the residual energy to estimate the capture point. Thisestimate is then used to update the weights of the RadialBasis Function. After approximately 500 learning trials, theRadial Basis Function converges and the residual energyerror is significantly reduced in further trials.
ReferencesJ. Rebula, F. Canas, J. Pratt, and A. Goswami. Learning
Capture Points for Bipedal Push Recovery. Humanoids 2007S. Kajita, F. Kanehiro, K. Kaneko, K. Yokoi, and H.
Hirukawa. The 3d linear inverted pendulum mode: a simplemodeling for a biped walking pattern generation. Proceed-ings of the IEEE/RSJ International Conference on IntelligentRobots and Systems, pages 239-246, 2001.
Jerry E. Pratt and Sergey V. Drakunov. Derivation andapplication of a conserved orbital energy for the invertedpendulum bipedal walking model. Proceedings of the IEEEInternational Conference on Robotics and Automation, pages4653-4660, 2007.
Jerry E. Pratt, John Carff, Sergey Drakunov, and Am-barish Goswami. Capture point: A step toward humanoidpush recovery. Proceedings of the IEEE-RAS InternationalConference on Humanoid Robots, 2006.
Russ Tedrake, Teresa Weirui Zhang, and H.SebastianSeung. Stochastic policy gradient reinforcement learning ona simple 3d biped. Proceedings of the IEEE InternationalConference on Intelligent Robots and Systems, pages 2849-2854, 2004.
2008 IEEE International Conference onRobotics and AutomationPasadena, CA, USA, May 19-23, 2008
978-1-4244-1647-9/08/$25.00 ©2008 IEEE. 1774