Research ArticleMEDICAL ROBOTS

Human-in-the-loop optimization of hip assistance with a soft exosuit during walking

See allHide authors and affiliations

Science Robotics  28 Feb 2018:
Vol. 3, Issue 15, eaar5438
DOI: 10.1126/scirobotics.aar5438
  • Fig. 1 Experimental setup for HIL Bayesian optimization.

    Bayesian optimization was used to adjust the control parameters of an assistive device to minimize the metabolic cost of walking. The metabolic rate was estimated from respiratory measurements and used to compute a posterior distribution of metabolic rate with respect to the free control parameters. The posterior was initially generated by evaluating six prefixed control parameters. Given the posterior at the current iteration, the control parameters with maximum EI were chosen and applied to the wearable device. This process was repeated until convergence. During this process, the configured force profiles were delivered through a soft hip exosuit with a tethered actuation system.

  • Fig. 2 Soft exosuit and assistive hip force profile.

    (A) The hip soft exosuit. A hip extension moment was generated by pulling the inner cable to create a tension between two anchor points. (B) Parameterization of hip force profile. The hip force profile was chosen to be a combination of two parameterized sinusoidal curves joined at the peak. Peak force was set to 30% of body weight, and onset timing was fixed to the time of maximum hip flexion. Peak and offset timing were actively adjusted by the optimization to determine the shape of the force profile as a function of gait percentage. Shaded purple and blue bars represent the range of peak and offset timing, respectively. (C) Examples of feasible hip force profiles.

  • Fig. 3 Experimental results.

    (A) The net metabolic rate for each condition. Optimal: Minimum mean value of the posterior distribution (metabolic landscape). Validation: Metabolic rate of 5-min walking with optimized assistance. No-suit: Metabolic rate of 5-min walking with a regular pair of pants. Bars are means, error bars are SEMs, and asterisks denote statistical significance. (B) Feasible parameter region and optimal timing values for all participants. Optimal timings were varied across participants, and three participants shared the same optimal timings at the latest peak and offset timing. (C) Optimal assistive force profiles for participants 3, 4, and 6. Dashed and solid lines are reference and measured forces normalized by body mass, averaged across 10 strides during the last minute of the validation condition. The maximum hip flexion event was used to initialize the gait cycle in this study.

  • Fig. 4 Participant-specific metabolic landscape and probability of improvement landscape.

    (A to C) Metabolic landscapes (the mean of the metabolic cost posterior distribution with respect to peak and offset timing) for participants 3, 4, and 6. Diamonds indicate the locations of participant-specific optimal timings. (D to F) Probability of improvement landscapes (capturing the probability of reducing metabolic cost beyond the identified optimal) for participants 3, 4, and 6.

Supplementary Materials

  • robotics.sciencemag.org/cgi/content/full/3/15/eaar5438/DC1

    Fig. S1. Illustration of 1D Gaussian process.

    Fig. S2. Simulation results on the number of iterations needed for the optimization.

    Fig. S3. Convergence analysis.

    Fig. S4. Optimized hip extension force profiles for all participants.

    Fig. S5. Experimental setup.

    Fig. S6. Experimental protocol.

    Fig. S7. Structure of the waist belt component.

    Fig. S8. Structure of the thigh brace.

    Fig. S9. Pseudo-randomly sampled timings for the initialization of Bayesian optimization.

    Fig. S10. Optimization process.

    Table S1. Signal-to-noise ratio and variations of metabolic cost of pilot tests.

    Table S2. Onset timing.

    Table S3. Participant characteristics.

    Table S4. Metabolic rates, optimal timing, and convergence timing for each participant.

    Table S5. Quadratic approximation of metabolic landscape.

  • Supplementary Materials

    Supplementary Material for:

    Human-in-the-loop optimization of hip assistance with a soft exosuit during walking

    Ye Ding, Myunghee Kim, Scott Kuindersma,* Conor J. Walsh*

    *Corresponding author. Email: walsh{at}seas.harvard.edu (C.J.W.); scottk{at}seas.harvard.edu (S.K.)

    Published 28 February 2018, Sci. Robot. 3, eaar5438 (2018)
    DOI: 10.1126/scirobotics.aar5438

    This PDF file includes:

    • Fig. S1. Illustration of 1D Gaussian process.
    • Fig. S2. Simulation results on the number of iterations needed for the optimization.
    • Fig. S3. Convergence analysis.
    • Fig. S4. Optimized hip extension force profiles for all participants.
    • Fig. S5. Experimental setup.
    • Fig. S6. Experimental protocol.
    • Fig. S7. Structure of the waist belt component.
    • Fig. S8. Structure of the thigh brace.
    • Fig. S9. Pseudo-randomly sampled timings for the initialization of Bayesian optimization.
    • Fig. S10. Optimization process.
    • Table S1. Signal-to-noise ratio and variations of metabolic cost of pilot tests.
    • Table S2. Onset timing.
    • Table S3. Participant characteristics.
    • Table S4. Metabolic rates, optimal timing, and convergence timing for each participant.
    • Table S5. Quadratic approximation of metabolic landscape.

    Download PDF

    Files in this Data Supplement:

Stay Connected to Science Robotics

Navigate This Article