Toward predicting exoskeleton-assisted gait

Team Members

• Ava Lakmazaheri
• Russell Martin

Project Video

Project Files

Project scripts (Matlab) 

Full results figures for various cost function (PDF)

Resultant model motion for GRF+kinematics+min(resid)+min(activations)

Background

Lower-limb exoskeletons have been demonstrated to reduce metabolic cost of walking. Human-in-the-loop optimization—a method of estimating exoskeleton characteristics that maximally benefit a given performance metric while the user is walking—has resulted in the largest metabolic reductions to date (Franks et al., 2021; Poggensee & Collins, 2021), but is expensive and time-consuming to perform. Biomechanical simulations can facilitate faster and cheaper optimization of exoskeleton assistance and studies of exoskeleton-assisted gait.

Franks and Bianco et al. (2020) optimized simulated hip-knee-ankle exoskeleton assistance to maximally reduce metabolic cost of walking. The resulting assistance strategy was estimated to reduce metabolic cost by 69.0% in simulation, but when tested experimentally, reduced metabolic cost by only 25.9%. Discrepancies between simulated and experimental gait could be due to a number of modeling choices: the simulation did not account for kinematic adaptations to assistance or a number of objectives known to guide human movement, including comfort (which was identified by the exoskeleton user during the experimental session) and balance (which prohibits muscle activity being driven to near-zero, as was observed for some simulated muscles).

Research Question(s)

Expanding on the work of Franks and Bianco et al. (2020), our overarching question is: what objectives matter for accurate prediction of exoskeleton-assisted gait?

In order to set up a predictive simulation of exoskeleton-assisted gait, we first created a tracking simulation of experimental data. For this stage of our research, we sought to understand how an increasing number of objectives—of tracking contact, tracking kinematics, minimizing contribution of reserve actuators, and minimizing control effort—capture relevant characteristics of unassisted walking.

Methods

Overview

Our proposed methods consist of three nested optimization and inverse optimization loops (Figure 1). In Loop 1, OpenSim Moco is used to predict joint kinematics and muscle activations that minimize a given cost function for a scaled musculoskeletal model with simulated bilateral ankle exoskeletons. Joint kinematics and muscle activations from unassisted walking—estimated in Moco by tracking experimental ground reaction forces (GRFs) and kinematics while minimizing control effort—will be used as an initial guess for the predictive simulation. For each run, errors will be calculated by comparing predicted outputs to ground truth data, for both joint kinematics and muscle activations. Weights on each cost function term will be tuned until the error converges, completing Loop 2. In Loop 3, this process will be repeated with different cost function terms as drawn from relevant literature. 

Figure 1. Systems flowchart for proposed methods

We analyzed data of one participant from Poggensee & Collins (2021). This study included trials of walking in normal shoes and walking with an ankle exoskeleton providing either zero torque, human-in-the-loop optimized torque, or generic torque assistance. Motion capture, force plate, exoskeleton torque, and electromyography data were collected for fifteen participants.

Adapting and scaling a musculoskeletal model

The musculoskeletal model was adapted to match participant anthropometrics and to include the marker set used in Poggensee & Collins (2021). We used AddBiomechanics to scale a generalized 3D, 37 degree of freedom, 80 muscle musculoskeletal model (Rajagopal et al., 2016) and applied the resulting segment scales factors to a 2D, 9 degree of freedom, 18 muscle musculoskeletal model (Ong et al., 2016), since AddBiomechanics cannot currently scale models with a planar joint, as was used to relate the ground to the pelvis in the latter model. The Ong model was chosen to facilitate faster iterations through the proposed nested optimization and inverse optimization loops and was also used in Franks and Bianco et al. (2020). We added ideal reserve actuators at each degree of freedom as well as a constant contact force sphere and half-space to the calcaneus and metatarsalphalangeal joint to model contact between the foot and the ground.  

Obtaining ground-truth unassisted walking data

Ground truth kinematic data for unassisted walking was calculated by performing inverse kinematics in OpenSim using the scaled musculoskeletal model and experimental motion capture data. Kinematic results were fitlered at 6 Hz and validated against normative data from the literature (Kadaba et al., 1990). Ground reaction force data were extracted from force plate data and filtered at 50 Hz. 

Data was visually inspected and trimmed to capture one stride of the left leg (a 1.25 second time window) following the five-minute mark of a six-minute walking trial, in order to best reflect steady-state gait behavior.

Tracking unassisted walking with minimum control effort

To perform motion tracking, we leveraged OpenSim Moco, a software toolkit to solve optimal control problems using direct collocation (Dembia et al., 2020). We input the scaled musculoskeletal model, experimental ground reaction force data, and ground truth joint angle trajectories. Optimization goals were to track ground reaction force data, track unassisted joint angle trajectories, minimize the contribution of reserve actuators, and minimize the sum of muscle activations squared. We performed these steps sequentially, beginning with the two tracking goals and no penalty on reserve actuators and muscle activations, before introducing penalties on reserves, followed by a penalty on control effort.

Results and Discussion

GRF and Kinematic Tracking

We first sought to track GRFs and kinematics while allowing the residual forces and muscle activations to be unconstrained. This allowed us to confirm our problem is dynamically feasible before trying to optimize control to be less reliant on reserves and more efficient. As expected, when prioritizing GRF and kinematic tracking alone, we could achieve very good results. Ground reaction force root mean square error (RMSE) were less than 5% of peak GRF, indicating appropriate levels of simulation precision. Similarly, tracking for the hip and ankle joint had RMSE less than 2 degrees. Curiously, the knee tracking was poor, with an RMSE of 13 degrees. We think this may be caused by dynamic inconsistencies due to how we model contact with the ground.

Figure 2. Ground reaction forces of model (blue) and reference data (red) for one stride of the left leg.

Figure 3. Sagittal kinematics data of the left leg for model (blue) and reference data (red).

Residual Reduction

Reduction of residuals, through increasing the residual penalty, performed as expected by shifting the force generation from reserve actuators to muscles. One simple example of this is in the ankle plantarflexor muscles (Figure 4). After penalizing the use of reserves, the force generated by the gastroc and soleus are larger to compensate, especially around toe-off. This phenomena was present at other joints as well, but is most easily visualized for the ankles.

Figure 4. Muscle force after adding a penalty for use of residual actuators.

Muscle Activity Reduction

Last, we sought to reduce the activity of muscles when they were not needed to match the desired kinematics and kinetics. Figure 5 shows sample activations of ankle joint muscles from our walking simulation, with a goal of tracking kinematics and GRFs and minimizing residual use and muscle activation squared. Comparison between the simulation and the experimental data illustrates similarities in salient features between the two. For example, plantarflexor muscles are moderately active in early-mid stance, more active in late stance/toe-off, and least active during swing. The tibialis anterior is highly active at heelstrike, less active during stance, and moderately active during swing. These similarities illustrate that, even with only four terms in the cost function, we can approximate relevant features in a Moco tracking simulation that might not be available from experimental data. Such similarities also suggest that our tracking simulation is well-posed, indicating we may be able to use these findings to inform our predictive simulation work. An animation of the walking model is shown in Figure 6.

Figure 5. Muscle activity from simulation (top row) and from experimental data (bottom row). Activation is normalized to maximum voluntary contraction; reference experiment data from Arnold et al. (2013). Tib Ant denotes tibialis anterior.

Figure 6. Animation of gait using all four cost function terms (GRF tracking, kinematic tracking, minimize residuals, minimize muscle activation squared).

Limitations

Our tracking simulation could be improved by looking at more than one stride and enforcing symmetry goals. Our model seems to "throw" itself forward in the last few frames (takes a longer step) - this could be reconciled by enforcing a constraint to make the initial and final kinematic states identical.

All findings above are from a single stride from a single participant. Results may improve by adding more participants (by reducing the effect of noise) or may worsen (if we had been overfitting to this participant). By running more participants through this protocol, we can learn about what parts of our model are most useful.

Future Work

We plan to continue working on our proposed inverse optimization methods outlined in Figure 1. Our next step will be to modify the scaled musculoskeletal model to add simulated ankle exoskeletons via ideal actuators at each ankle joint with torque profiles and mass properties reflecting the device used in Poggensee & Collins (2021). We will then set up a predictive simulation in OpenSim Moco using the results from our final tracking optimization as an initial guess (Figure 1, Loop 1) before calculating error relative to the experimental data and tuning weights to minimize error (Figure 2, Loop 2). Finally, we will iterate on objective function terms to gain insights on what control objectives best reflect exoskeleton-assisted gait. With access to 15 participants walking in this experimental setup, each with a variety of exoskeleton assistance strategies, we can then assess the generalizability of our approach.

Acknowledgments

We thank Nick Bianco, Jon Stingel, Delaney Miller, Keenon Werling, Scott Delp, and Steve Collins for their guidance and technical assistance. We thank Katie Poggensee for the experimental data and Apoorva Rajagopal and Carmichael Ong for the musculoskeletal models.

References

Dembia, C. L., Bianco, N. A., Falisse, A., Hicks, J. L., & Delp, S. L. (2020). OpenSim Moco: Musculoskeletal optimal control. PLOS Computational Biology, 16(12), e1008493. https://doi.org/10.1371/journal.pcbi.1008493

Franks, P. W., Bianco, N. A., Bryan, G. M., Hicks, J. L., Delp, S. L., & Collins, S. H. (2020). Testing Simulated Assistance Strategies on a Hip-Knee-Ankle Exoskeleton: A Case Study. 2020 8th IEEE RAS/EMBS International Conference for Biomedical Robotics and Biomechatronics (BioRob), 700–707. https://doi.org/10.1109/BioRob49111.2020.9224345

Franks, P. W., Bryan, G. M., Martin, R. M., Reyes, R., Lakmazaheri, A. C., & Collins, S. H. (2021). Comparing optimized exoskeleton assistance of the hip, knee, and ankle in single and multi-joint configurations. Wearable Technologies, 2, E16. https://doi.org/10.1017/wtc.2021.14

Kadaba, M. P., Ramakrishnan, H. K., & Wootten, M. E. (1990). Measurement of lower extremity kinematics during level walking. Journal of Orthopaedic Research, 8(3), 383–392. https://doi.org/10.1002/jor.1100080310

Ong, C. F., Hicks, J. L., & Delp, S. L. (2016). Simulation-Based Design for Wearable Robotic Systems: An Optimization Framework for Enhancing a Standing Long Jump. IEEE Transactions on Biomedical Engineering, 63(5), 894–903. https://doi.org/10.1109/TBME.2015.2463077

Poggensee, K. L., & Collins, S. H. (2021). How adaptation, training, and customization contribute to benefits from exoskeleton assistance. Science Robotics, 6(58), 1–13. https://doi.org/10.1126/scirobotics.abf1078

Rajagopal, A., Dembia, C. L., DeMers, M. S., Delp, D. D., Hicks, J. L., & Delp, S. L. (2016). Full-Body Musculoskeletal Model for Muscle-Driven Simulation of Human Gait. IEEE Transactions on Biomedical Engineering, 63(10), 2068–2079. https://doi.org/10.1109/TBME.2016.2586891

Arnold, E. M., Hamner, S. R., Seth, A., Millard, M., & Delp, S. L. (2013). How muscle fiber lengths and velocities affect muscle force generation as humans walk and run at different speeds. Journal of Experimental Biology, 216(11), 2150-2160. https://doi.org/10.1242/jeb.075697