Real-time Control of Walking for Biomechanical Models of Bipeds

This page presents the project Real-time Control of Walking for Biomechanical Models of Bipeds started by PhilippeBeaudoin under the supervision of MichielVanDePanne.

Description

Human motion can be modeled kinematically or dynamically. Dynamic models potentially offer the deepest understanding of human motion given that they model the forces and torques that give rise to the motion. Dynamic biomechanical models further detail by modeling the motion in terms of muscle activations which drive muscles connected to tendons and bones [1,2]. An understanding of how activities such as robust control of walking can be achieved with musculotendon models is a significant step in enabling patients with a variety of neuro-motor deficits to possibly walk again [3]. With the recent development of robust strategies for physically-simulated walking with torque-based actuators [4], it should now be feasible to develop similar control strategies for musculo-tendon based models. While the joint-specific feedback-error learning approach of [4] needs to be replaced by a muscle-specific approach, the recent work of [5] provide a very promising path to addressing exactly this problem. Models of reaction delays and tripping reflexes can also be added. This would be the first real-time, closed-loop biomechanical simulation of walking of its kind.

References

• [1] "Musculotendon Simulation for Hand Animation", SIGGRAPH 2008
• [2] "Simulation of Walking", 2006 http://nmbl.stanford.edu/publications/pdf/Anderson2006.pdf
• [3] "Optimal Patterns Synthesis Approach for Knee Motion under Functional Electrical Stimulation" http://hal-lirmm.ccsd.cnrs.fr/lirmm-00193720/en/
• [4] "SIMBICON: Simple Biped Locomotion Control", SIGGRAPH 2007 http://www.cs.ubc.ca/~van/papers/Simbicon.htm
• [5] "Recurrent Cerebellar Loops Simplify Adaptive Control of Redundant and Nonlinear Motor Systems", Neural Computation, 2006

Papers Read

• "Simulation of Walking", 2006 http://nmbl.stanford.edu/publications/pdf/Anderson2006.pdf Comments
• Delp, S.L., Anderson, F.C., Arnold, A.S., Loan, P., Habib, A., John, C.T., Guendelman, E., Thelan, D.G. OpenSim: Open-source software to create and analyze dynamic simulations of movement. IEEE Transactions on Biomedical Engineering , vol 55, pp 1940-1950, 2007. http://nmbl.stanford.edu/publications/pdf/Delp2007.pdf Comments

Papers to read

• Bill Sellers. Muscle Mechanics. http://homepage.mac.com/wis/Personal/lectures/musculoskeletal/MuscleMechanics.pdf
• Regina Nuzzo. Computational Biomechanics: Making Strides Toward Patient Care. Biomedical Computation Review 2006. http://biomedicalcomputationreview.org/3/1/3.pdf
• Rachel Weinstein, Eran Guendelman, and Ron Fedkiw. Impulse Based Control of Joints and Muscles. TVCG 2006. http://physbam.stanford.edu/~fedkiw/papers/stanford2006-08.pdf
• Winnie Tsang, Karan Singh, Eugene Fiume. Helping Hand: An Anatomically Accurate Inverse Dynamics Solution For Unconstrained Hand Motion. SCA 2005. http://www.dgp.toronto.edu/~karan/pdf/HandPaper_SCA.pdf
• Sung-Hee Lee, Demetri Terzopoulos. Heads up!: biomechanical modeling and neuromuscular control of the neck, SIGGRAPH 2006. http://www.cs.ucla.edu/~sunghee/research/neck/siggraph06.pdf
• Helping Hand: An anatomically accurate inverse dynamics solution for unconstrained hand motion, SCA 2005
• Automatic Determination of Facial Muscle Activations from Sparse Motion Capture Marker Data", SIGGRAPH 2005, ACM TOG 24, 417-425 (2005).
• Heads up!: biomechanical modeling and neuromuscular control of the neck SH Lee, D Terzopoulos - Proceedings of ACM SIGGRAPH 2006, 2006 - portal.acm.org
• Impulse-based control of joints and muscles R Weinstein, E Guendelman, R Fedkiw - IEEE Transactions on Visualization and Computer Graphics, 2008 -ieeexplore.ieee.org

Links

• Project To Do List
• Project Results

Discussion

Discovering muscle activation from torque-based simbicon

One idea I have for this would be to linearize the torque-->muscle function around the steady state of a torque-based simbicon controller. Then we can simply run our torque-based controller (given the current character state) and use our linear relationship to get muscle activations. Since linearization is around steady-state, it can be precomputed. Some remaining questions are:

• What to do outside steady state?
• How far can we take our linear model?
• Do we have to linearize on-the-fly when the character state is far from what we know?
• Do we want to skip the "torque" stage and directly get muscle activations (or forces?) from the target angles (i.e. some kind of muscle-PD-controller)?

Also note that the "character state" is higher dimensional here than in torque-based Simbicon. This is because muscles are generally modeled as dynamical systems (using an ODE), so the "past state" of the muscle has an influence on the activation-->force function. Linearizing around steady-state eschew this problem, but it can appear again during perturbations. Can we build a simple model of the force-->activation function? Then we still have the problem of torque-->force, but this requires a lower-dimensional character state (only angles? or angular velocities too?).

Clean-up the following, Michiel's email I think that it may ultimately be interesting to apply the feedback error learning ideas to the muscles. The cool thing is that you could keep the feedback controller based upon regular PD controllers, while gradually training a muscle-based feedforward term. I'm still not sure what to do about antagonistic muscles. But the above scheme would not yet use muscles to achieve feedback-based corrections. Just an idea. Maybe it is still better to first get a full muscle-based solution going for one of the joints.

-- PhilippeBeaudoin - 04 Feb 2009