Multi-expert learning of adaptive legged locomotion

Achieving versatile robot locomotion requires motor skills that can adapt to previously unseen situations. We propose a multi-expert learning architecture (MELA) that learns to generate adaptive skills from a group of representative expert skills. During training, MELA is first initialized by a distinct set of pretrained experts, each in a separate deep neural network (DNN). Then, by learning the combination of these DNNs using a gating neural network (GNN), MELA can acquire more specialized experts and transitional skills across various locomotion modes. During runtime, MELA constantly blends multiple DNNs and dynamically synthesizes a new DNN to produce adaptive behaviors in response to changing situations. This approach leverages the advantages of trained expert skills and the fast online synthesis of adaptive policies to generate responsive motor skills during the changing tasks. Using one unified MELA framework, we demonstrated successful multiskill locomotion on a real quadruped robot that performed coherent trotting, steering, and fall recovery autonomously and showed the merit of multi-expert learning generating behaviors that can adapt to unseen scenarios.

Fall Inducing Movable Platform (FIMP) for Overground Trips and Slips

Background: The study of falls and fall prevention/intervention devices requires the recording of true fallsincidence. However, true falls are rare, random, and difficult to collect in real world settings. A system capableof producing falls in an ecologically valid manner will be very helpful in collecting the data necessary toadvance our understanding of the neuro and musculoskeletal mechanisms underpinning real-world falls events.Methods: A fall inducing movable platform (FIMP) was designed to arrest or accelerate a subject's ankle toinduce a trip or slip. The ankle was arrested posteriorly with an electromagnetic brake and acceleratedanteriorly with a motor. A power spring was connected in series between the ankle and the brake/motor toallow freedom of movement (system transparency) when a fall is not being induced. A gait phase detectionalgorithm was also created to enable precise activation of the fall inducing mechanisms. Statistical ParametricMapping (SPM1D) and one-way repeated measure ANOVA were used to evaluate the ability of the FIMP toinduce a trip or slip.Results: During FIMP induced trips, the brake activates at the terminal swing or mid swing gait phase toinduce the lowering or skipping strategies, respectively. For the lowering strategy, the characteristic leg loweringand subsequent contralateral leg swing was seen in all subjects. Likewise, for the skipping strategy, all subjectsskipped forward on the perturbed leg.Slip was induced by FIMP by using a motor to impart unwanted forward acceleration to the ankle with thehelp of friction-reducing ground sliding sheets. Joint stiffening was observed during the slips, and subjectsuniversally adopted the surfing strategy after the initial slip.Conclusion: The results indicate that FIMP can induce ecologically valid falls under controlled laboratoryconditions. The use of SPM1D in conjunction with FIMP allows for the time varying statistical quanticationof trip and slip reactive kinematics events. With future research, fall recovery anomalies in subjects can nowalso be systematically evaluated through the assessment of other neuromuscular variables such as joint forces,muscle activation and muscle forces.

Fall Inducing Movable Platform (FIMP) for Overground Trips and Slips

The study of falls and any related fall prevention/intervention device requires the recording of true falls incidence. However, true falls are rare, random and difficult to collect. Therefore, a system that can perturb falls in an ecologically valid and repeatedly manner will greatly benefit the understanding of the neuromuscular mechanisms underpinning real-world falls events. A fall inducing movable platform (FIMP) was designed to arrest and accelerate the subject's ankle to induce trip via a brake and slip via a motor respectively. A gait phase detection algorithm was also created to allow the timely activation of the fall mechanisms to induce different recovery actions. Statistical Parametric Mapping (SPM1D) and two sample t-test were used to evaluate the transparency of the platform before it was used to induce falls. Thereafter, SPM1D and one-way repeated measure ANOVA were used assess the effectiveness of FIMP in inducing realistic falls. Walking with the FIMP's fall mechanisms attached on the ankle (SW) was found to be similar to normal walking (NW), except for a slight increase in ankle flexion during the swing phase. However, the magnitude of change would be considered negligible when compared to the changes in joint angles during the trips and slips of interest. During the FIMP induced trips, the brake activates at the terminal-swing and mid-swing gait phase to induce the lowering and skipping strategies respectively. The characteristic leg lowering and the subsequent contralateral leg swing was seen in all subjects for the lowering strategy. Likewise, for skipping strategy, all subjects skipped forward on the perturbed leg. On the other hand, slip was induced by FIMP using the motor to impart unwanted forward acceleration to the ankle with the help of friction-reducing ground sliding sheets. Joints stiffening was observed during slips, and subjects adopt the \textit{surfing} strategy after the initial slip. Results indicate that FIMP can induce reliable and ecologically valid falls repeatedly under simulated experimental conditions. The usage of SPM1D with FIMP allows the creation of the first ever quantifiable trip and slip reactive kinematics comparison. Effects of fall recovery anomalies can now be easily identified.