Gait Trajectory and Event Prediction from State Estimation for Exoskeletons During Gait.

A real-time method is proposed to obtain a single, consistent probabilistic model to predict future joint angles, velocities, accelerations and jerks, together with the timing for the initial contact, foot flat, heel off and toe off events. In a training phase, a probabilistic principal component model is learned from normal walking, which is used in the online phase for state estimation and prediction. This is validated for normal walking and walking with an exoskeleton. Without exoskeleton, both joint trajectories and gait events are predicted without bias. With exoskeleton, the trajectory prediction is unbiased, but event prediction is slightly biased with a maximum of 33 ms for the toe off event. Performance is compared with predictions based on only the population mean. Without exoskeleton, estimation errors are 5 to 30% lower with our method. With exoskeleton, trajectory prediction errors are up to 20% lower, but gait event prediction errors only improve for foot flat (30%) and are worse for other events (30%-50%). The ability to predict future joint trajectories and gait events offers opportunities to design exoskeleton controllers which anticipate these trajectories and events, allowing better tracking control and smoother, accurately timed transitions between different control modes.

A probabilistic method to estimate gait kinetics in the absence of ground reaction force measurements.

Human joint torques during gait are usually computed using inverse dynamics. This method requires a skeletal model, kinematics and measured ground reaction forces and moments (GRFM). Measuring GRFM is however only possible in a controlled environment. This paper introduces a probabilistic method based on probabilistic principal component analysis to estimate the joint torques for healthy gait without measured GRFM. A gait dataset of 23 subjects was obtained containing kinematics, measured GRFM and joint torques from inverse dynamics in order to obtain a probabilistic model. This model was then used to estimate the joint torques of other subjects without measured GRFM. Only kinematics, a skeletal model and timing of gait events are needed. Estimation only takes 0.28 ms per time instant. Using cross-validation, the resulting root mean square estimation errors for the lower-limb joint torques are found to be approximately 0.1 Nm/kg, which is 6-18% of the range of the ground truth joint torques. Estimated joint torque and GRFM errors are up to two times smaller than model-based state-of-the-art methods. Model-free artificial neural networks can achieve lower errors than our method, but are less repeatable, do not contain uncertainty information on the estimates and are difficult to use in situations which are not in the learning set. In contrast, our method performs well in a new situation where the walking speed is higher than in the learning dataset. The method can for example be used to estimate the kinetics during overground walking without force plates, during treadmill walking without (separate) force plates and during ambulatory measurements.

Model-based control for exoskeletons with series elastic actuators evaluated on sit-to-stand movements.

BACKGROUND: Currently, control of exoskeletons in rehabilitation focuses on imposing desired trajectories to promote relearning of motions. Furthermore, assistance is often provided by imposing these desired trajectories using impedance controllers. However, lower-limb exoskeletons are also a promising solution for mobility problems of individuals in daily life. To develop an assistive exoskeleton which allows the user to be autonomous, i.e. in control of his motions, remains a challenge. This paper presents a model-based control method to tackle this challenge. METHODS: The model-based control method utilizes a dynamic model of the exoskeleton to compensate for its own dynamics. After this compensation of the exoskeleton dynamics, the exoskeleton can provide a desired assistance to the user. While dynamic models of exoskeletons used in the literature focus on gravity compensation only, the need for modelling and monitoring of the ground contact impedes their widespread use. The control strategy proposed here relies on modelling of the full exoskeleton dynamics and of the contacts with the environment. A modelling strategy and general control scheme are introduced. RESULTS: Validation of the control method on 15 non-disabled adults performing sit-to-stand motions shows that muscle effort and joint torques are similar in the conditions with dynamically compensated exoskeleton and without exoskeleton. The condition with exoskeleton in which the compensating controller was not active showed a significant increase in human joint torques and muscle effort at the knee and hip. Motor saturation occurred during the assisted condition, which limited the assistance the exoskeleton could deliver. CONCLUSIONS: This work presents the modelling steps and controller design to compensate the exoskeleton dynamics. The validation seems to indicate that the presented model-based controller is able to compensate the exoskeleton.

Subject-Exoskeleton Contact Model Calibration Leads to Accurate Interaction Force Predictions.

Knowledge of human-exoskeleton interaction forces is crucial to assess user comfort and effectiveness of the interaction. The subject-exoskeleton collaborative movement and its interaction forces can be predicted in silico using computational modeling techniques. We developed an optimal control framework that consisted of three phases. First, the foot-ground (Phase A) and the subject-exoskeleton (Phase B) contact models were calibrated using three experimental sit-to-stand trials. Then, the collaborative movement and the subject-exoskeleton interaction forces, of six different sit-to-stand trials were predicted (Phase C). The results show that the contact models were able to reproduce experimental kinematics of calibration trials (mean root mean square differences - RMSD - coordinates ≤ 1.1° and velocities ≤ 6.8°/s), ground reaction forces (mean RMSD≤ 22.9 N), as well as the interaction forces at the pelvis, thigh, and shank (mean RMSD ≤ 5.4 N). Phase C could predict the collaborative movements of prediction trials (mean RMSD coordinates ≤ 3.5° and velocities ≤ 15.0°/s), and their subject-exoskeleton interaction forces (mean RMSD ≤ 13.1° N). In conclusion, this optimal control framework could be used while designing exoskeletons to have in silico knowledge of new optimal movements and their interaction forces.

Real-Time Gait Event Detection Based on Kinematic Data Coupled to a Biomechanical Model.

Real-time detection of multiple stance events, more specifically initial contact (IC), foot flat (FF), heel off (HO), and toe off (TO), could greatly benefit neurorobotic (NR) and neuroprosthetic (NP) control. Three real-time threshold-based algorithms have been developed, detecting the aforementioned events based on kinematic data in combination with a biomechanical model. Data from seven subjects walking at three speeds on an instrumented treadmill were used to validate the presented algorithms, accumulating to a total of 558 steps. The reference for the gait events was obtained using marker and force plate data. All algorithms had excellent precision and no false positives were observed. Timing delays of the presented algorithms were similar to current state-of-the-art algorithms for the detection of IC and TO, whereas smaller delays were achieved for the detection of FF. Our results indicate that, based on their high precision and low delays, these algorithms can be used for the control of an NR/NP, with the exception of the HO event. Kinematic data is used in most NR/NP control schemes and is thus available at no additional cost, resulting in a minimal computational burden. The presented methods can also be applied for screening pathological gait or gait analysis in general in/outside of the laboratory.