Dynamic Walking
Submitted by mike on Fri, 2009-08-07 04:17

Mike Stilman's History of Dynamic Walking

As with all histories this is probably quite biased. Please let me know if we are missing important references and papers or simply add them.

All Japanese humanoids are Dynamic walkers. If you hear someone claim that they are static please correct them. Static walking makes the assumption that the robot is not moving and it is very slow. The walks do not look weird because they are static. They might look awkward because they are not fully optimized or because we still don't have enough redundancy in the joints to generate perfectly fluid motion. The main distinction in present day walking is not between Static and Dynamic but rather between Dynamic and Passive Dynamic.

  • Dynamic walking is any type of walk that places some guarantees on dynamic stability.
  • Passive Dynamic walking takes advantage of the natural system dynamics such as momentum, springs and the cyclic nature of the walk to make it natural and significantly lower the amount of energy expended.

Note that Japanese humanoids (and Golem Krang) typically use very high ratio gears (such as Harmonic Drives) which force one to ignore any concept of passivity and rather just force the robot to do whatever it needs to do. Chris Atkeson has refferred to this approach as "imposing our will on the robot" and it has been the standard approach to controlling robots since the early days of factory automation. Chris is a proponent of Passive Dynamics - I'm on the fence.


Origins of Biped Robot Locomotion

1966-69
1970
  • Miomir Vukobratovic (Yugoslavia/Serbia) proposed the Zero Moment Point as the stability criterion for walking animals and robots.
    Vukboratovic, Zero Moment Paper 1
    Vukboratovic, Additional Info in 2004

    In contrast to the Center of Mass criterion used in statics, ZMP takes into account the motor torques at the joints. There is some contention about Vukobratovic's invention since it was apparently known in biophysics long before that. One thing is clear, Vokobratovic brought it to the attention of roboticists.

1970s
  • Waseda gets really into fluid power and works a lot on hydraulics, pneumatics and artificial muscles for actuating robots. Kato develops the WABOT-1, the first humanoid robot. This is still a static walker.
  • McGhee (USA) at Ohio State develops numerous controllers for various legged walkers.
1980
  • Waseda is still using 10 seconds per step on walking but starts making claims about quasi-dynamic walks.
1982
1984
1986
  • Honda (Japan) starts development of a walking humanoid robot.
  • Marc Raibert writes a massive paper to ACM describing various walking and hopping machines. He also details the last five years of accomplishments in the CMU Leg Lab. Cool paper!
    http://portal.acm.org/citation.cfm?id=5948.5950
1989
  • Takanishi (Waseda) publishes an automated method for generating a walking pattern based on linearizing the ZMP equations and solving for the inputs in the frequency domain. The solution isthen converted back to the time domain and error is estimated on the full model. The process is iterated until error is minimized and a stable trajectory is achieved.
    http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=637905
1990
1991


Modern ZMP Walking

1994
  • JSK, organized by H. Inoue and M. Inaba at the University of Tokyo is created to study humanoid robotics. Particular goals are walking, vision, manipulation and high level reasoning for humanoid robots.
1995
1996
  • Honda announces P2 a full body humanoid robot in development since 1986. The robot achieves a desired ZMP with some additional compensation for ground disturbance. It is able to walk, scale stairs and handle minor disturbances.
1998
  • METI (Japanese Ministry of Economy) finances a 5 year Humanoid Robotics Prject (HRP). AIST (formerly MEL), Honda and Kawada Industries are involved. HRP-1 a retrofitted Honda P3 is shown off to perform various tasks such as operating a forklift on a rainy day and in-house service robotics.
2000
  • First International Conference on Humanoid Robotics (Humanoids '00)
  • S. Kagami and K. Nishiwaki (JSK) propose a discrete time method for quickly generating walking trajectories by incremental optimization in the time domain. (Can we find the original paper on this? The later journal version incorporates work form 2001/2002).
2001
  • Nishiwaki proposes interpolating pre-computed motions to achieve a space of possible displacements for real-time control of humanoid walking. The trajectories now allow specified velocities at the boundaris such that they may be connected. The approach is implemented on the JSK / Kawada robot H6.
    http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=933260

  • AIST (in Tsukuba, formerly MEL) organizes the Humanoid Robotics Group under H. Hirukawa, lead by K. Yokoi with very notable walking experts: S. Kajita and K. Kaneko.
  • AIST (in Odaiba/Tokyo) creates the Digital Human Laboratory / Research Center lead by T. Kanade with more notable walking experts from JSK: K. Nishiwaki and S. Kagami.
2002
  • K. Nishiwaki (JSK/AIST) shows an online pattern generation method for walking which computes entirely new motions online by iterative optimization. We now see the "overlapped trajectory" planning structure. Updated trajectories are generated every few milliseconds that connect to previous ones. The method is demonstrated on JSK/Kawada H7.
    http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1041675
    also here:
    http://www.jsk.t.u-tokyo.ac.jp/~report/proceeding02/iros02/nishiwaki_fas...

  • S. Kajita (AIST) presents a distinct 3D-LIP model based online walking method with a closed form solution. This is based on the concept of a linear controller that generates a repulsive force field on a single mass. The approach is demonstrated on HRP-2L (A block torso mounted on the legs of future HRP-2).
    http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1013335

  • Kawada and AIST announce the completion of HRP-2. It is made available to a number of institutes in Japan including the two AIST locations and JSK.
2003
2004
2004-09 By 2004, planar biped walking using ZMP was largely considered a solved problem. Preview Control was accepted by many researchers as the standard solution with other methods available for cases when they appeared more relevant. Research in ZMP walking started to focus on: (1) Improving stability with ground irregularities and external forces. (2) Walking in environments with obstacles. (3) Using walking in conjunction with manipulation to perform higher level tasks. Here are some samples of this kind of research:

Improving Stability:

Environments with Obstacles:

Higher Level Tasks:
2009
  • AIST and Kawada present the HRP-3 a feminine android that walks, talks and expresses emotion.


Modern Passive Dynamic Walking

To be completed...
AttachmentSize
WL-3-1969.jpg233.32 KB
WAP-3-1971.jpg423.03 KB
3D_hopper.jpeg109.21 KB
Vukobra2.pdf395.77 KB
Vukobra1.pdf3.69 MB
saika.gif8.43 KB
hrp1-header.jpg34.6 KB
robot_model_hrp_1_dago7.jpg343.93 KB
»