
Engineering and Technology
Personalizing exoskeleton assistance while walking in the real world
P. Slade, M. J. Kochenderfer, et al.
This study showcases rapid exoskeleton optimization in real-world settings, achieving remarkable improvements in walking speed and energy efficiency. Conducted by Patrick Slade, Mykel J. Kochenderfer, Scott L. Delp, and Steven H. Collins, it highlights the effectiveness of portable ankle exoskeletons using wearable sensors, outperforming traditional lab methods significantly.
~3 min • Beginner • English
Introduction
The study addresses how to personalize ankle exoskeleton assistance under naturalistic, real-world conditions without relying on laboratory-only measures such as indirect calorimetry during long bouts of steady treadmill walking. Although exoskeletons can improve walking speed and reduce energetic cost in labs, translating these benefits to daily life is challenging because (1) the performance criterion used for personalization (metabolic rate) traditionally requires expensive equipment, (2) real-world walking occurs in numerous short bouts at varying speeds, and (3) assistive devices must be portable and cost-effective. The authors hypothesize that information embedded in human movement, measurable with wearable sensors, can be used in a data-driven optimization framework to quickly identify beneficial assistance parameters during everyday walking. The purpose is to develop and validate a portable exoskeleton and a real-time, data-driven optimization method that personalizes assistance across realistic walking bouts and speeds, thereby making exoskeleton benefits practical outside the laboratory.
Literature Review
Prior work shows exoskeletons can increase walking speed and reduce metabolic energy in laboratory settings, particularly when assistance is individualized via human-in-the-loop optimization. However, personalization has typically required measurement of metabolic rate and extended, steady-state treadmill trials, which are impractical for clinics or real-world use. Alternatives like musculoskeletal modeling could estimate performance but are computationally intensive and require subject-specific tuning. The authors leverage a previously collected large dataset (~3,600 assistance conditions) combining lab measurements with wearable sensor data, using it to enable a data-driven approach that predicts which assistance patterns reduce metabolic cost without direct metabolic measurements.
Methodology
Device: A portable, untethered ankle exoskeleton was developed with a low mass (~1.2 kg per ankle). A brushless motor with a custom drum-and-rope transmission applied assistive torque about the ankle. Electronics on-board sensed motion and executed real-time control and optimization on a microcontroller. Torque control used classical feedback and iterative learning, achieving <1% peak torque tracking error. The system could apply high peak torque (reported up to 6.4 N·m at 1.5 m s⁻¹) while maintaining motor temperature well below thermal limits during extended use. A 0.3 kg battery powered at least 30 min of walking.
Data-driven optimization model: Using a prior dataset in which participants walked with ~3,600 different exoskeleton assistance conditions, the team trained a logistic regression classifier to predict which of two control laws would yield lower metabolic energy. Inputs included ankle angle and ankle velocity signals segmented by gait cycle and the torque parameters of the two control laws being compared. For each pair of control laws, the model computed a pair coefficient via weighted differences in segmented motion features and, via a logistic function, the probability that one law was superior to the other. Each control law received a score by summing probabilities across all its pairwise comparisons, producing a ranking.
Optimization loop: During an optimization session, a participant experienced a sequence of k control laws (parameter sets defining torque profiles). After collecting steps for each law, the model compared all pairs, ranked the laws, and an optimizer updated the estimate of optimal parameters and selected a new set of laws to evaluate. This cycle repeated until convergence criteria were met. Computation ran in real time on the exoskeleton’s microcontroller.
Short-bout, variable-speed data handling: Because everyday walking comprises many short bouts (with 90% <100 steps) and variable speeds, the method used metadata from each step. Tests showed control-law comparisons were reliable with as few as ~44 continuous steps, enabling opportunistic data capture from most bouts. The system binned data by walking speed, associated each step with a bin, and performed data-driven ranking within bins when sufficient data accumulated. Rankings then updated nominal parameter estimates across speed bins.
Speed estimation and speed-adaptive control: Step-by-step speed was estimated from stride period using a model (RMSE ≈ 0.06 m·s⁻¹). A speed-adaptive controller interpolated between previously optimized parameters to estimate per-step optimal assistance as walking speed changed.
Experimental protocols: 1) Laboratory validation with a tethered ankle exoskeleton compared data-driven optimization against metabolic-rate-based optimization, quantifying time to convergence and metabolic outcomes (n=9). 2) Naturalistic outdoor optimization: participants walked for about one hour on a public course/sidewalk, following randomized audio prompts eliciting a distribution of bout speeds and durations that matched real-world statistics. Optimization proceeded over many short bouts, with convergence tracked. 3) Field validation: in separate outdoor bouts with randomized conditions, ground-truth metabolic cost and speed were measured to quantify benefits of Real-world Optimized assistance versus Normal Shoes and a Generic assistance baseline. Additional tests included treadmill trials with sinusoidal speed variation (0.75–1.75 m·s⁻¹) to assess speed-adaptive control.
Key Findings
- Data-driven optimization matched the effectiveness of laboratory metabolic optimization but identified optimal parameters approximately 4× faster (n=9); the resulting metabolic cost with data-driven optimized parameters was within 5% of metabolic-optimized parameters.
- Personalized, Real-world Optimized assistance achieved during about one hour of naturalistic walking increased self-selected speed by 9 ± 4% (ANOVA, n=10, P=0.031) and reduced cost of transport by 17 ± 5% (ANOVA, n=10, P=0.039) relative to Normal Shoes. The speed increase (~0.12 m·s⁻¹) is meaningful, and energy savings approximate removing a 9.2 kg backpack.
- On a treadmill at 1.5 m·s⁻¹, exoskeleton assistance reduced metabolic energy consumption by 23 ± 8% compared with normal shoes.
- Individual participants exhibited distinct optimized parameter sets, differing from Generic assistance.
- Speed-adaptive assistance reduced metabolic cost more than fixed Generic assistance when treadmill speed varied sinusoidally between 0.75 and 1.75 m·s⁻¹ (n=3).
- The portable device provided high, accurately controlled torque with low tracking error; maximum assistance did not overheat the motor, and battery supported at least 30 min of walking.
Discussion
The findings demonstrate that human movement signals captured by wearable sensors contain sufficient information to guide personalization of exoskeleton assistance without direct metabolic measurements. The data-driven, pairwise-comparison approach effectively identified beneficial parameter sets under fragmented, variable-speed walking typical of daily life, overcoming longstanding barriers of laboratory dependency. Personalized assistance produced meaningful improvements in speed and energy economy outdoors, aligning with prior lab results and confirming that translating benefits to the real world is feasible. Speed-adaptive control further improved performance as walking speed changed, highlighting the value of per-step adaptation. The convergence behavior indicated that longer optimization durations could yield even better parameter estimates. Pilot observations suggest benefits extend to other conditions (inclines, load carriage, stairs), and user feedback indicated the untethered system was comfortable and usable. Overall, the study supports deploying real-time, on-body personalization to enhance mobility for both clinical populations and high-performance users.
Conclusion
This work introduces a portable ankle exoskeleton and a real-time, data-driven optimization framework that personalizes assistance during naturalistic walking. The approach achieves laboratory-comparable benefits in a fraction of the time and produces significant improvements in outdoor walking speed and energy economy. Contributions include: (1) a classifier-based human-in-the-loop optimization method using wearable sensor kinematics, (2) a short-bout, speed-binned data strategy enabling real-world personalization, (3) a speed-adaptive controller for per-step parameter interpolation, and (4) a compact device capable of accurate, high-torque assistance. Future research could extend optimization time for improved convergence, validate across broader populations and environments, integrate additional sensing modalities, and explore multi-joint assistance and longer-duration power systems to support all-day use.
Limitations
- Optimization during the one-hour outdoor sessions did not reach steady state; additional time may improve parameter estimates and benefits.
- Sample sizes were modest in several tests (e.g., n=3 for some speed-adaptive evaluations; n=9–10 for primary comparisons), which may limit generalizability.
- Battery life supported at least 30 minutes of walking, potentially constraining longer field use without recharging or larger batteries.
- Real-world testing occurred on a specific outdoor course with guided prompts; performance in unconstrained daily life and diverse terrains requires further validation.
- Assistance and analyses focused on the ankle joint; benefits for different tasks or user populations (e.g., clinical groups) need dedicated studies.
Related Publications
Explore these studies to deepen your understanding of the subject.