Updated:May 25, 2017
Dynamic Walking 2017

Meeting participants got the video-link password in email. [Option: to keep the video folder open all the time, click here.].

Fixing problems. If you find broken links or other issues, please email us us. If your entries need corrections or additions, use the upload page. Use your 3-digit number below when naming files, e.g., 476-Stefany-Singh.jpg (or .mov or .mp4 or .pdf or .txt for short abstracts and other comments.corrections).

[ Apologies if we garbled your short abstract. You can upload a new one to the dropbox link. Please keep words and sentences short. Don't use unexpanded acronyms. Take account the variety of expertises, and lack thereof, in your audience.]


341
- (5 minute talk)
341 Murthy Arelekatti (Arelekatti, Winter)
We present an inverse dynamics approach to select optimal mechanical components to enable the knee moment required for able-bodied kinematics of level ground walking. We also propose a novel architecture of the knee mechanism with timely engagement and disengagement of the chosen mechanical components to achieve reliable stance and swing phase control.
abstract; video

401
The closed-loop motor control in locomotion (0.5 minute talk)
401 Berat Denizdurduran (Denizdurduran, Gewaltig)
Because locomotion is basically underactuated and also has hybrid dynamics, control design is difficult. And there is no working synthetic neural controller either. We focus on the dynamics and biomechanical to aid thinking about the neural controller design.
video

400
Inverse Optimal Control for a Simple Stepping Task
400 John Rebula (Rebula, Mason, Schaal, Righetti)
What is being minimized in legged locomotion? We use trajectory optimization. From each set of constraints on the motion and weights on various proposed cost components we predict a preference. We do inverse optimization to infer the weights of the optimization from observed motion.
abstract

399
- (0.5 minute talk)
399 Wolfgang Rampeltshammer (-)
I am interested in control strategies, and the principles that generate them, for wearable robots/exoskeletons/protheses for walking. And I want to hear recent insights from human locomotion studies.
video

187
- (0.5 minute talk)
187 Sean Mason (-)
I am interested in control methods that allow robots to walk over rough terrain. I also want to learn about new actuation and sensor technologies.
video

507
How many Cinxia live in the land Islands, and why? (30 minute talk)
507 Saskya van Nouhuys (-)
Melitaea cinxia is a species of butterfly. Here, in land, there is a decades-long, millions-of-euros, 100s of people, 1000s of km^2, study of cinxia, along with the plants cinxia caterpillars eat, and the insects that eat cinxia. It is one of the biggest ecological studies in the world. Ill tell you about a wasp that is a parasite of cinxia. It can find virtually all the cinxia caterpillars. But, luckily, it only kills some of them. What keeps it in check?
abstract; video

402
Finding agile movements of robots with underactuation and unilateral constraints: The Butterfly robot case studies (12 minute talk)
402 Anton Shiryaev (Surov, Shiryaev, Freidovich, Gusev, Paramonov)
The talk is aimed at a discussion of principles for organization of a search for agile movements for mechanical systems with dynamic constraints. The focus will be on the example of planning movements for the Butterfly robot, which is the benchmark robotic system that has been conceptually designed and approached by many researchers, working with non-prehensile manipulations, within the last two decades. The system is inspired by motions of circus jugglers. It is a one-link curvy robotic hand that is expected to perform a dynamic manipulation of an object, which can freely roll on its surface. Rolling of an object, as rolling a foot during walking, represents a unilateral constraint, while sliding and non-sliding as well as under-actuation, for both cases: of manipulation and walking, are dynamical constraints inherently present in the systems description. Scalable arguments and steps in planning feasible behaviors of the benchmark system and controlling them are presented. They are complemented and illustrated by simulations and experiments, performed on the robotic set-ups in our lab.
abstract; video

508
The amazing Aland Islands (30 minute talk)
508 Graham Robbins (-)
I'm English, an archeologist, and had a son with a Swedish-speaking Finn. So, I ended up working here, in the museum which you will visit. Aland used to be underwater, squashed by the ice. When it popped up, 10,000 years ago, people came. They ate fish and killed seals. Since then it keeps popping, about 5 mm each year. So, there has been more and more land for Russia and Sweden to fight each other for. That's part of why there are 100s of sunken ships around here.

387
- (0.5 minute talk)
387 Wendy Boehm (-)
I am new to control theory and simulations. I want to use our models of human walking and standing to explore and quantify the stability of control algorithms.
video

393
- (0.5 minute talk)
393 Siyuan Feng (-)
I am interested in not-fragile control through contacts. I am from the model-based optimization camp, so I would like to extrapolate from that knowledge. The goal is to embrace the world.
video

396
- (0.5 minute talk)
396 Songyan Xin (Romeo Orsolino, Yapeng Shi)
I am interested in model-based optimization applied to humanoid locomotion. Our lab uses these ideas on a torque-controlled humanoid robot. I am eager to learn from others.
video

347
Mina V2: Exoskeleton for Paraplegic Mobility (5 minute talk)
347 Jeremy Gines (Peter Neuhaus, Jeremy Gines, Tyson Cobb, Travis Craig, Jesper Smith, and Robert Griffin)
Is toe off torque useful for paraplegic mobility? We tested walking with and without toe off torque. We saw greater mobility and stability when using toe off torque.
abstract; video

397
- (0.5 minute talk)
397 Ryan Elandt (-)
I am brand new to this field, I want to learn the diverse control ideas for walking humanoids. I also am excited to see the creative modern processes used by others.
video

392
- (0.5 minute talk)
392 Yves Nazon II (-)
I am excited to learn about the control techniques used for legged robots and wearable robots that interact with humans during gait.
video

190
- (2 minute talk)
190 Qingguo Li (-)
I want to learn more about how and why external wearable devices affect normal gait and about the interaction between the user and a device.
video

192
- (0.5 minute talk)
192 William Zhiren Peng (-)
I want to know about the energetics of bipedal humans and bipedal battery-powered robots. I am a co-author of Carlotta Mummolo's abstract: "Contact-dependent balance stability of biped systems."
video

365
Disturbance Observer Based HZD Control of Biped Walking and Slip Recovery (2 minute talk)
365 Yoshitaka Abe (Abe, Chen, Trkov, Yi, and Katsura)
Slip recovery of a biped walking robot is challenging, even in 2D. Human-inspired trajectory generation and DOB (Disturbance OBserver) based control are implemented. We conducted nonslip, 2D walking experiment. We simulated slip, and its recovery, in 2D walking.
abstract; video

386
High-Order Contact-Implicit Trajectory Optimization (0.5 minute talk)
386 Scott Kuindersma (Zachary Manchester, Scott Kuindersma)
Most contact-implicit (i.e., non-hybrid) trajectory optimization algorithms use first-order integrators that create a practical tradeoff between problem size and trajectory accuracy (which, in turn, affects tracking performance). We will be presenting recent work exploring the use of variational integrators for rigid-body systems to develop new contact-implicit trajectory optimization algorithms with high-order integration accuracy.
video

337
Sensitivity analysis of the balance stability region in legged mechanisms (2 minute talk)
337 Carlos Isaac Gonzalez Bolivar (Gonzalez, Mummolo, Kim)
How sensitive is the balance stability of a legged system to variations in parameters, e.g., joint torque limits and size of the base of support? Within optimization-based dynamics, we did sensitivity analyses, both analytically and numerically. There are saturation points for both the joint torque limits and for the support-base size. They are co-related.
abstract; video

184
Contact-dependent balance stability of biped systems (5 minute talk)
184 Joo Kim (Mummolo, Peng, Gonzalez, Park, Kim)
How do contacts alter the balance stability of biped systems? The complete set of balanced and falling states of biped systems in a specified (multi-)contact configuration, such as single and double supports, are identified and evaluated. The stability regions of a biped robot and a set of human subjects are compared, along with their walking characteristics.
abstract; video

281
Energy efficient control of a 1D hopper through tunable damping (5 minute talk)
281 Uluc Saranli (Secer, Saranli)
We seek a low-power method to control energy for running. We propose to embed a SLIP (spring mass) model, with tunable damping coefficient, into a serially actuated compliant leg. Experiments show that this is feasible, and also improves efficiency and accuracy.
abstract; video

285
Cascade Non-Linear State Estimation for Humanoid Robot Locomotion (3 minute talk)
285 Stylianos Piperakis (Piperakis and Trahanias)
We have a cascade estimation scheme, consisting of two Extended Kalman Filters (EKF), for acquiring low-drift state estimates of important quantities commonly used in control loops for either humanoid robot balance or locomotion.
abstract; video

287
Fast, Sampling-Based Kinodynamic Bipedal Locomotion Planning with Moving Obstacles (3 minute talk)
287 Junhyeok Ahn (Ahn, Campbell, Kim, Sentis)
Can we do kino-dynamic planning using a sampling-based method for a high dimensional humanoid robot? We used the "rapid exploration of random trees" (RRT). By storing key walking information in our nodes we were able to plan over hundred steps. Using this, we expect our humanoid can go through a maze-like room filled with both static and moving obstacles.
abstract; video

286
How (de)coupled are Minitaur limbs? (5 minute talk)
286 Andrew Pace (Andrew Pace, Sam Burden)
Does the robot Minitaur, from Ghost Robotics, inertially decouple the limbs and the body. We will measure both body and limb acceleration and velocity at impact. Does the body have significant instantaneous velocity change. Current results for one leg, in certain configurations, suggest that the body has minimal velocity change just after ground contact.
abstract; video

265
A new model for legged locomotion. (5 minute talk)
265 Vikas Bhandawat (Biswas, Rao, Bhandawat)
We want the simplest physics-based model for legged locomotion which captures the key features across all speeds, all animals, and different numbers of legs. We use a combination of kinematic measurements and modeling. In an attempt to fit measurements of fruit flies, we found that no existing locomotion model fits the data. So, we made up a new model. This new model has features which might make it of general utility for legged locomotion.
abstract; video

379
- (5 minute talk)
379 Amy Wu (-)
Does complexity (in terms of modelling and hardware) necessarily lead to usefulness or understanding of human locomotion? Much can be answered with simple simulation models, and likewise it should be possible to test biomechanical principles with low-cost materials and methods (e.g. the DRAPA Challenge!), scaling size and complexity to the scientific question of interest. This reach for low cost robots aims to bridge the gap between researchers and society by improving accessibility to robotics research and promoting scientific understanding of locomotion behaviors.
video

195
#248 Learning to Hop using Guided Policy Search on Real Robot Hardware / #234 Linking Mechanics and Learning (0.5 minute talk)
195 Felix Jakob Grimminger (-)
I want to make high-performance legged robots. So, I would like to learn more about the dynamic locomotion. We have been working on the mechatronic design of a small, electric, torque-controlled robot using brushless motors and field oriented control. My colleagues will present our initial results using a first prototype leg in a linear test stand. (#248 / #234)
video

364
Bio-inspired Control Framework for Legged Locomotion (4 minute talk)
364 Yapeng Shi (Shi,Cai,Guo,Jiang,Wang,Xin)
We want both compliant locomotion, and precise movements, in a challenging terrain. So, we made a bio-inspired control framework. It is based on the Equilibrium Point Hypothesis. Our simulations and experiments verify the feasibility of the Equlibrium Point method.
abstract; video

380
Human Feedback Control to Maintain Trajectories of Task-Relevant Variables During Sit-to-Stand Motion (0.5 minute talk)
380 Patrick Holmes (Holmes, Danforth, Moore, Fu, Vasudevan)
We characterize the feedback laws that moving humans use when responding to perturbations. We carried out an eleven-person perturbative Sit-To-Stand (STS) experiment, built individualized inverted pendulum models from motion capture data, and compared models with both feedforward and feedback control to the observed human responses to perturbation. We also performed a motor equivalence analysis which showed that during STS, humans respond to perturbations in a way that maintains the nominal trajectories of the x-position of both their center of mass and head.
abstract; video

381
- (0.5 minute talk)
381 In-Hwan Kim (Kim, Yi)
The dynamic-walking method for biped robots needs continuous stepping for balancing even in constant position. An active wheel applied on a foot (see recent Boston Dynamics videos) is proposed to solve this problem. This Leg-Wheel combined structure provides biped robots with the ability to walk in rough terrains, climb stairs and get around with wheels in flat terrains.
video

372
Flexible spine modeling for quadruped robot (0.5 minute talk)
372 Shounak Bhattacharya (Bhattacharya, Ghosal)
We investigate shapes of flexible spine in quadruped motion. Optimization finds the best shape. Simulated quadruped motion is more natural looking when we us an optimized flexible spine.
abstract; video

382
Estimation of Human Ankle Impedance During Late Stance Phase of Walking (8 minute talk)
382 Elliott Rouse (Shorter, Rouse)
We can now measire ankle mechanical impedance during walking. This view goes beyond the traditional descriptions of joint kinetics and kinematics. However, it is hard to measure impedence during the push-off phase of walking, when the majority of mechanical work is added. Preliminary results find the impedance of the human ankle during the push-off phase of walking, and compare this to previous results concerning the mid-stance and swing phase of gait.
abstract; video

374
Design of a Bilateral Lower Limb Exoskeleton Emulator (5 minute talk)
374 Patrick Franks (Franks, Bryan, Klein, Collins)
We propose the design of a bilateral lower limb exoskeleton emulator. We will use human-in-the-loop optimization to investigate assistance at the hip, knee, and ankle during walking and running.
abstract; video

339
Scaling of inertial delays in terrestrial mammals (5 minute talk)
339 Sayed Naseel Mohamed Thangal (Mohamed Thangal, Donelan)
How does inertia contribute to an animal's response time and how does this scale with animal size. We use computational models of a simple pendulum and of an inverted pendulum. We quantify inertial delay as the time taken to execute specified tasks. We found that inertial delays scale slower than delays predicted by geometric similarity. But inertial delays are still long enough to dominate over sensorimotor delays, at least for small movements in larger animals .
abstract; video

258
Ramp perturbation tests are too simple to identify a realistic controller in human standing balance (3 minute talk)
258 Anne Koelewijn (Wang, Koelewijn, Van den Bogert)
Whatexperiments can identify the control system for human standing balance? Perturbation experiments and indirect system identification, fit the closed loop model responses to human data. Ramp perturbation experiments are fitted easily but do not produce a unique system identification when the model includes muscle dynamics. Random perturbation experiments produce locally optimal controller parameters which depend on the initial guess. Further improvement in numerical methods is needed to overcome these issues.
abstract; video

175
- (0.5 minute talk)
175 Robert Griffin (Neuhaus, Gines, Cobb, Craig, Smith, Griffin)
I want to hear the latest in optimization-based dynamic planning for humanoid robots. I also want to hear how the hardware field has progressed in the last year. Lastly, I hope to learn about new biological observations in walking so I might apply them to robot control.
video

322
Human-in-the-loop optimization strategies to improve both balance in steady-state walking and active engagement in post-stroke gait rehabilitation (5 minute talk)
322 Katherine Poggensee (Poggensee, Sreenath, Collins)
This past year, we have shown the success of a human-in-the-loop optimization strategy. The mechanisms behind the success of the algorithm are not entirely understood. With this new study, we hope to identify the relative contributions of general assistance strategies, customization, and facilitated learning.
abstract; video

203
Passive inter-joint gait assisting exoskeleton suit (3 minute talk)
203 Jun-tian Zhang (Zhang, Li)
We experimentally compare three device configurations, using only a passive approach: a) ankle-knee joint assistance, b) hip assistance and c) combined assistance. I expect the ankle-knee configuration to provide the most assistance, followed by the hip only configuration, and finally the ankle only configuration will likely give the least assistance.
abstract; video

378
Learning an Efficient Walking Trot using Variable Impedance (5 minute talk)
378 Elco Heijmink (Heijmink, Radulescu, Semini)
Can an appropriate learning methodology improve the efficiency of a quadrupedal walk, with minimal need of expert knowledge and a short learning time? We propose a motor learning methodology that can improve the efficiency of a trotting gait by learning of an impedance modulation strategy. The method shows a significant improvement within a couple of minutes, both in simulation and on hardware.
abstract; video

299
Viscosity-based Height Reflex for Quadrupedal Locomotion on Rough Terrain (5 minute talk)
299 Michele Focchi (Focchi, Featherstone, Orsolino, Caldwell, Semini)
Problem: leg mobility-loss when swinging down from very high steps. My proposal: height reflex replanning strategy which "spreads" the swing motion in a smart way onto the other stance legs (e.g. making the trunk squat). Results: I made a 80 kg quadruped robot step down from a pallet that was 26cm high, which is 58% of the retractable leg range.
abstract; video

296
Joint velocity measurement using low-cost high bandwidth MEMS gyroscopes. (4 minute talk)
296 Jesper Smith (Jesper Smith, Georg Wiedebach and Sylvain Bertrand)
To accurately predict the instantaneous capture point we need to accurately measure joint velocities, supressing large spikes due to backlash. We use low cost MEMS sensors connected to an EtherCAT network allowing up to 500 sensors at 1kHz. The initial prototype showed good velocity estimation of the spine joint of Atlas.
abstract; video

312
For commonly employed muscle-tendon-complex models, predictions of human metabolic energy expenditure, predictions of human metabolic energy expenditure are substaintially lower than measured values at the single joint level (10 minute talk)
312 Koen Kasper Lemaire (Lemaire, van der Zee, Reuvers and van Soest)
Do commonly employed muscle tendon complex models correctly predict mechanical behavior and metabolic energy expenditure? We measured and modelled subjects' torque about the knee joint, and metabolic energy expenditure, during a kinematically contrained knee flexion-extension task. We found that for both a Hill and a Huxley muscle models, mechanical behavior was predicted well, but energy use was way underestimated.
abstract; video

310
Predicting human-structure interaction with a stable walking controller: synchrony and entrainment on the London Millennium Bridge (0 minute talk)
310 Varun Joshi (Joshi, Srinivasan)
What caused pedestrians on the London Millennium Bridge to synchronize with each other and eventually shake the bridge? Using simple models of walking, we tried to predict human gait on the bridge. We used energy minimization and stability as independent predictors. Both hypotheses predict the shaking bridge. The human need for stability can produce synchronization.
abstract; video

273
Towards limits of energy efficiency in legged locomotion (5 minute talk)
273 Fabio Giardina (Giardina, Iida)
We made simple mechanical models, both theoretical and physical, which can explain the limits of energy efficiency in legged hopping locomotion. We made four models and robots, some to minimize the cost of transport and explore actuation methods, and some to minimize complexity in both morphology and control. We are eager to find quantifiable conditions for which legs are superior to wheels.
abstract; video

305
Is maximum effort acceleration limited by leg force generation capabilities? (15 minute talk)
305 Alison Sheets-Singer (Sheets-Singer, Vinet)
A previous claim is that an athlete's sprint speed is limited by his/her ability to produce larger leg forces. Do athletes actually maximize leg force and power production during a maximum effort linear acceleration? Eight male athletes did 10m accelerations from a standing position, with and without 10% body mass added to the torso. With and without the additional torso mass, during maximal-effort accelerations there were no differences between the peak resultant ground reaction force, nor of center of mass power.
abstract; video

247
Individual muscle contributions to the position of the centre of pressure (5 minute talk)
247 Raphael Dumas (Raphael Dumas, Florent Moissenet)
We aim to estimate the individual muscle contribution to the position of the Center of Pressure (CoP). We use a 3D lower-limb musculoskeletal model together with simple optimisation methods (inverse dynamics, pseudo-inverse). This allow to assess which muscles mainly contribute to the anterior- posterior and medial-lateral displacement of the CoP during the stance.
abstract; video

246
How can we achieve versatile and robust robotic walking? The concept of Divergent Component of Motion as possible candidate (8 minute talk)
246 Johannes Englsberger (Englsberger)
How to achieve versatile and robust walking? I choose the Divergent Component of Motion as tool. It results in promising simulation results, yet practical issues have yet to be solved/discussed to achieve fully satisfactory performance.
abstract; video

388
- (30 minute talk)
388 Art Kuo (-)
I wasnt proposing to give a talk, but would be happy to fill in if needed. We do have a recent and truly marvelous optimization result, which this text box is too small to contain.

249
Hybrid averaging shows that within-stance symmetry helps mitigate coupling interactions between degrees of freedom in a sagittal 2DOF monoped and 3DOF biped (15 minute talk)
249 Avik De (De, Koditschek)
Raibert's planar hopping controller has three decoupled 'parts' (e.g, the vertical hopping controller ignores fore-aft motion, etc.). We had success with similar decoupling strategies on a tailed biped and on a quadruped. Why and when we can ignore coupling interactions? How can de-coupling be leveraged to synthesize useful running behavior? The key new insight is that time-reversal symmetry of the Hamiltonian components of within-stance dynamics (in a simple 2D and 3D model) allows for favorable coupling-cross-term cancellations in the *averaged* return map linearization.
abstract; video

375
Creating controlled perturbations to study stumble recovery strategies (3 minute talk)
375 Guan Rong (Tan, Emanuel, Collins)
We hope to design control algorithms and prostheses that can help transfemoral amputees better apply stumble recovery strategies. To do so, we designed a mechanism that can apply desired-force-profile sperturbations to various body parts. This allows us to study reflex recovery strategies in able bodied and amputee subjects. From this data, we hope to be able to better understand how to design control strategies and hardware which can implement the oberved strategies.
abstract; video

385
Learning indirect optimal control for dynamic motion planning with RRT (10 minute talk)
385 Wouter Wolfslag (Wolfslag, Bharatheesha, Moerland, Wisse)
We want to plan on the fly, so our robots can avoid obstacles. Sampling-based motion planning is good for that, but needs lots of online computation. We have an offline learning approach, based on indirect optimal control. We've tested the swing-up problem, and hope for success with tougher problems.
abstract; video

183
Influence of Voluntary Intervention on Gait Entrainment (8 minute talk)
183 Meghan Huber (Ochoa, Huber, Hogan)
My talk is joint with Julieth Ochoa (#267). Humans synchronize to rhythmic mechanical, visual, and auditory stimuli. I want to know how and why. What about locomotor control causes this entrainment, especially when the stimulus is not mechanical. Does feedback from multiple sources play a big role in the neural control of gait?
video

373
Locomotion Control of Passive Dynamic Walkers based on Internal Energy (2 minute talk)
373 Andrea Calanca (Piccinelli, Calanca)
We seek ROBUST limit-cycle stability of simple models, such as the compass or slip models. Our simple, continuous and physically-inspired control law is based on energy shaping. In simulation this controller enhances stability: continuous walking is observed even with small perturbations. Video: https://youtu.be/qG1WGr6o7Es
abstract; video

390
Approximate explicit model predictive control for push recovery using mixed-integer convex optimization (0.5 minute talk)
390 Twan Koolen (Koolen, Marcucci, Manuelli, Koolen, Tedrake)
How can we get humanoid robots to recover from pushes on rough terrain? We believe recent advances in mixed-integer optimization can help to determine where the robot's feet and hands should be placed in order to recover from a push. We hoped that results from these mixed-integer optimizations could be precomputed exactly, but are now leaning towards using approximation techniques.
video

327
Characterizing the relationship between step length asymmetry and metabolic rate during locomotion in post-stroke individuals (5 minute talk)
327 Thu Nguyen (Nguyen, Jackson, Aucie, Torres-Oviedo, Collins)
Do individuals who suffer from chronic stroke choose to walk with a step length asymmetry to minimize their energy consumption? We ask subjects to modulate their step length asymmetry using visual feedback and motion capture, and we measure their energy consumption. We hypothesize that their energy cost, to walk symmetrically or with an exaggerated asymmetry, will be higher than their nominal asymmetry.
abstract; video

204
Comparing the Energetic Effects of Different Energy Harvesting Profiles (0.5 minute talk)
204 Michael Shepertycky (Shepertycky, Burton, Dickson, Liu, Li)
We predict that a harvesting profile that mimics the user's natural net knee moment could better assist the user, and therefore lead to a larger decrease in the metabolic cost of walking, than a constant resistance profile. To test this, we will conduct human treadmill walking experiments.
abstract; video

216
A passive spring-mass model with rolling contact and leg masses to investigate the ballistics of swing dynamics. (6 minute talk)
216 Will Charles (Charles, Ransing, Sazonov, Friswell)
Relative to the rigid inverted pendulum model for walking, a spring-loaded mass model can improve the predicted of ground reaction forces. We add a swing leg to the model, like on the classic passive dynamic compass walker, and get a stable limit cycle. We look at the effect of leg stiffness on the stability of gait, as measured by the Basin of Attraction. What insight can this model give to the mechanics of gait?
abstract; video

284
Using Equilibria and Virtual Holonomic Constraints to Generate Families of Walking Gaits (8 minute talk)
284 Nelson Rosa Jr. (Rosa, Lynch)
Can we find periodic motions for biped robots from a single motion template, including when all joints are passive? We use the equilibria of a biped as a motion template, and continuously deform them into walking gaits. This way, we made gaits for various planar point- and curved-foot walkers, including 2D point-foot models of MARLO and Atlas. We have promising results for extensions to 3D bipeds.
abstract; video

298
Interactive walking pattern generator for mechanically coupled bipedal agents (10 minute talk)
298 Jessica Lanini (Lanini, Ajallooeian, Ijspeert)
People walking together often synchronize their foot-fall patterns. We can be model this synchronization using a coupled SLIP (spring mass) model. Could this idea generate walking pattern of humanoids? By reducing the complexity of the SLIP, using an AMP, we predict coordinations seen in humans.
abstract; video

292
Modeling and exploring elderly walking with neuromechanical simulations (12 minute talk)
292 Seungmoon Song (Song, Geyer)
We adapt our neuromechanical simulation model to study why elderly people use more energy to walk and prefer to walk slower. So, we modified our young-person model to represent healthy elderly people by applying skeletal, muscular and neural changes. We then analyze the resulting walking simulations. The decline of gait performance is mainly due to reduced muscle strength.
abstract; video

377
Tik-Tok: a high-performance legged locomotion platform (7 minute talk)
377 Jason Cortell (Cortell)
The electric biped/legged locomotion platform Tik-Tok aims to approach human-level locomotion performance in energy effectiveness, some kinds of agility, and robust balance. Energy use is minimized through use of high-efficiency brushless motors and chain drives; high peak power capability at the joints allows rapid repositioning of the feet for robust maintenance of balance. This new robot is intended to traverse normal human built environments safely and reliably (sidewalks, floors, and eventually stairs) with minimal power use. We report on our construction progress.
abstract; video

271
Experimental Validation of the Lower Leg Trajectory Error, an Optimization Metric for Prosthetic Feet (0.5 minute talk)
271 Victor Prost (Prost, Olesnavage, Winter, Major)
We have a new metric for prosthetic foot design: Lower Leg Trajectory Error (LLTE). LLTE connects the stiffness and geometry of a prosthetic foot with its biomechanical performance. LLTE will let us design single-component compliant feet that are optimized for various activities; can be cheap, through injection molding or with customizable 3D printing; and are tunable for specific patients' body weight, size, and level of activity. (See also the Winter talk)
abstract; video

395
Mapping the Influence of Spatiotemporal Asymmetries on Energetic Cost and Reactive Balance during Walking (12 minute talk)
395 James Finley (Finley, Liu, Sanchez)
We use voluntary manipulation of spatiotemporal asymmetry to map the relationship between asymmetry, energetic cost, and reactive balance control during walking on a split-belt treadmill. Our results refute the idea that symmetry is energetically optimal when walking on a split-belt treadmill and point to balance control as a potential alternative objective influencing how people adapt to walking in an asymmetric environment.
abstract; video

398
Frog musculo-robotics (30 minute talk)
398 Chris Richards (Richards, Eberhard, Porro, Collings)
Using frogs as a model, we aim to understand how musculo-skeletal anatomy influences limb function. To test this, we manipulate the anatomical properties of computational and/or bio-robotic limb models which are actuated by living frog muscle tissue. We expect that small changes in muscle moment arms will strongly impact jump performance by influencing the degree of active muscle lengthening prior to takeoff.
video

241
Inertial lever as a mechanical principle of human walking (2 minute talk)
241 Peter Jakubik (Jakubik)
Just what is the exact principle of human locomotion? Do we need a new principle? I noticed unexpected dynamics in a simple 2-link pendulum model of human walking. So, I introduced the notion of an Inertial Lever. I claim that the Inertial Lever is the principle of human locomotion. I showed analytically that forward propulsion is a natural result of the Inertial Lever, because internal lever forces act like muscle forces.
abstract; video

240
A singularity-tolerant inverse kinematics including joint position and velocity limitations (5 minute talk)
240 Salman Faraji (Faraji, Ijspeert)
Singularities and joint limits are difficult to handle, but might improve efficiency. We have a flexible nonlinear optimization, with safe operation regions and good tracking. It runs in real-time, approaches singularities safely and comes out of them without delay.
abstract; video

177
Proposed Bilateral Lower Limb Exoskeleton Emulator Design (0.5 minute talk)
177 Gwen Bryan (Patrick Franks, Stefan Klein, Steve Collins)
We have designed a bilateral lower limb exoskeleton emulator to provide torque to the ankle, knee and hip joints of both legs. We will use human-in-the-loop optimization to reduce metabolic cost of human locomotion. During this process, we will investigate assistance strategies with a variety of speeds, grades, and loads.
abstract; video

383
An Open-source Robotic Ankle Prosthesis (5 minute talk)
383 Alejandro Azocar (Azocar, Rouse)
In this talk, we present the design of a robust, inexpensive, and open-source prosthetic ankle that permits testing of control algorithms in the lab, community, and at home. The prosthesis is meant to provide a hardware platform for controls researchers to test their algorithms beyond simulations or bipedal robots, and reduce the barriers to entry for performing clinically beneficial research.
video

384
Real-time feedback modalities for training neuromuscular coordination during walking and standing following stroke (12 minute talk)
384 Kreg Gruben (Gruben, Boehm)
We aim to understand the interaction of control and mechanics that enable balance during human bipedal tasks. We quantify and model intact and disrupted human behavior, and then use that information to design and build therapeutic devices. The goal of those devices is to restore the balance component of disrupted control in order to improve safety and efficacy of human bipedalism.
abstract; video

330
Contact-dependent balance stability of biped systems (8 minute talk)
330 Carlotta Mummolo (Mummolo, Peng, Gonzalez, Park, Kim)
How do contacts alter the balance stability of biped systems? The complete set of balanced and falling states of biped systems in a specified (multi-)contact configuration, such as single and double supports, are identified and evaluated. We compared the stability regions and walking trajectories of a biped robot and human subjects.
abstract; video

254
Design of a Passive Prosthetic Foot with a Tension Energy Recovery (TER) System (4 minute talk)
254 Ben Morgan (Ben Morgan, Liam Cotton, Fawzi Belblidia, Dr. Rajesh Ransing & Dr. Afshin Tarat)
Our new energy storage and return system assists push-off on a passive prosthetic foot. I am checking if this design works better than other existent devices. My experiment will integrate into ANSYS simulation software. The ISO 10328 standard ensures safety, and topological optimisation helps refine the design. We predict that we can help drive the COM forward and reduce unhealthy sideways forces. We want to restore patient's gait to their healthy natural gait.
abstract; video

255
Steps in model-based trajectory searching as a tool for biped design (4 minute talk)
255 Torleif Anstensrud (Paramonov, Anstensrud, S_tre, Shiriaev)
We're exploring the design and development of a simple three-link walking machine. The physical prototype is designed based on an existing dynamic model, where the virtual-holonomic-constraints approach is used to search for energy-efficient gaits. Simulations find a rich variety of feasible low-torque walking gaits. We hope our experiments will validate the approach.
abstract; video

229
Optimal control describes quadrupedal walking in dogs (4 minute talk)
229 Delyle T Polet (Polet, Bertram)
Do domestic dogs walk in a way that minimizes work? We used trajectory optimization of a simple quadrupedal model with realistic morphological inputs, and set the cost function to work-based cost of transport. The optimal gait was similar to the actual gaits used by dogs with similar morphology.
abstract; video

391
Mechanical power output in periodic motions: the frame of reference matters (10 minute talk)
391 Lotte Lintmeijer (Lintmeijer, Hofmijster, Soest)
Commercially available instruments supposedly measure a rower's mechanical power output. They use the product of the moment of the handle force (at the oar pin) and the oar angular velocity. However, this approach is wrong. It is based on the incorrect assumption that a boat-fixed frame of reference constitutes an inertial frame of reference. Both computer simulations and experimental evaluation indicate that the veridical (actual) rower's power output is at least 10% larger than the power output 'measured' by the commercial proxy.
video

357
Three Uses for Springs in Extension in Legged Locomotion (15 minute talk)
357 Aaron M. Johnson (Johnson)
Compliant robot legs are typically used in compression. However, sometimes it may be possible, and advantageous, to use the legs in tension. We show three such uses: terrain adaptation, energy storage, and shape change. The general strategies are shown with robot experiments.
abstract; video

358
The Effect of Spine Morphology on the Motions and Energetics of Quadrupedal Robots (0.5 minute talk)
358 William Yang (Yang, Yesilevskiy, Remy)
We examine and quantify the effects of spine morphology on the motion and energetics of quadrupedal robots. We use optimization to compare the most energetically optimal motion of each model. We will test various spine characteristics, such as the number of segments and increased stiffness in specific spinal areas.
abstract; video

282
Better safe than sorry: stochastic optimal control of gait predicts larger foot clearance (8 minute talk)
282 Anne Koelewijn (Koelewijn, Van den Bogert)
Does a stochastic human model predict more-realistic movements than a deterministic human model? We solve trajectory optimization problems on several noisy episodes of movement instead of a single deterministic episode. The stochastic model predicts more-realistic foot clearance.
abstract; video

267
"Influence of Voluntary Intervention on Gait Entrainment" (0 minute talk)
267 Julieth Ochoa (Ochoa; Huber; Hogan)
When doing a distractor task, unimpaired and impaired subjects' gaits tend to entrain to periodic ankle torque pulses. We assess the influence of voluntary intervention on entrainment. Subjects' gait adaptation, when explicitly instructed to 'entrain', differed considerably compared to when distracted. So, active cognitive engagement of the Central Nervous System may evoke additional neural processes that interfere with "natural" walking. This is a joint talk with Meghan Huber (#183).
abstract; video

269
Human Feedback Control to Maintain Trajectories of Task-Relevant Variables During Sit-to-Stand Motion (10 minute talk)
269 Shannon Maureen Danforth (Holmes, Danforth, Moore, Fu, Vasudevan)
What feedback laws dictate human reponse to dynamic perturbations? We did an 11-person perturbative Sit-To-Stand (STS) experiment, built individualized inverted pendulum models from motion capture data, and compared models with both feedforward and feedback control to what perturbed people do. We found that during STS, humans respond to perturbations in a way that maintains one coordinate of the nominal tractory of both the center of mass and of the head.
abstract; video

295
A First Step Towards Adapting the 3D Reflex Based Neuromuscular Gait Model for Gait Assistive Devices (10 minute talk)
295 Arvid Keemink (Brug, Keemink, van Asseldonk, van der Kooij)
Can the 3D Song/Geyer (3DSG) neuromuscular gait model reproduce human responses to force perturbations (i.e. forward/backward pushes) at the pelvis? We compared how human subjects and the model responded to these forces. The 3DSG model can't respond properly at the ankle, even after re-optimization with human kinematic data. So, we added extra control modules to create a human-like torque response at the ankle.
abstract; video

317
Safe Online Learning Using Barrier Functions (12 minute talk)
317 Nils Smit-Anseeuw (Smit-Anseeuw, Remy, Vasudevan)
We have a way to guarantee the safety of online learning. The method uses barrier certificates and Sums-of-Squares programming to create 'training wheels'. These can be added to controller rollouts to guarantee safe learning. The idea works with a simple car model and a simple hopping model. We're hoping it works for a simple walking model too.
abstract; video

300
Turning Maneuvers Strategy for Quadrupedal Bounding Gait (5 minute talk)
300 Romeo Orsolino (Orsolino, Focchi, Caldwell, Semini)
Online lanning of highly dynamic gaits of quadruped robots, using a whole-body model, is too computationally expensive to react to a disturbance or to a command of an external operator. Our approach couples the CoM (Center of Mass) dynamics with two typical stability criteria of legged locomotion: periodic limit cycles (for the forward speed) and ZMP (for the frontal plane). The result is a controller that allows us to have our robot HyQ perform omni-directional turning maneuvers while keeping the desired stable bounding gait.
abstract; video

304
A low-cost, actuated passive dynamic walker kit for accessible research and education (5 minute talk)
304 Matthew Robertson (Robertson, Paik, Ijspeert, Wu)
Bipedal robotics research is often limited by lack of access to expensive humanoid platforms. But, many of the principles of locomotion and control may not need such expensive machines. To fill the vacancy between 'serious' state-of-the-art walking robots and simplified walking 'toys', we built a low-cost, servo-actuated walking robot using readily available materials. We show that this walker can achieve dynamic powered walking using a simple control scheme. Wigh minimal intervention it can start walking too.
abstract; video

242
Inclusion of total angular momentum in performance criterion improves prediction of healthy gait (3 minute talk)
242 Wouter Aerts (W. Aerts, F. de Groote, I. Jonkers, J. Vander Sloten)
We made a multiple shooting method for predictive forward simulations of walking. Minimizing total angular momentum predicts observed healthy gait.
abstract; video

248
Learning to Hop using Guided Policy Search on Real Robot Hardware (10 minute talk)
248 Julian Viereck (Viereck, Grimmiger, Herzog, Righetti )
We try to find what kind of algorithms can learn very dynamic locomotion. We use Guided Policy Search [GPS, Levine et al. '15], which works on other learning problems. In simulation the GPS algorithm learned hopping motions. We are now testing the algorithm on hardware. We 3D printed a small robot leg. The robot is easily repaired and can "learn" over days, without manual resets.
abstract; video

336
The cause of energetic cost differences in walking and running: optimization modeling and speed-gravity experiments (10 minute talk)
336 John EA Bertram (Bertram, Hasaneini, Schroeder, Ruina)
Why is the energetic cost of walking highly sensitive to speed but only modestly sensitive to gravity- reduction while the cost of running is just the converse. A cost-of-transport optimization model replicates empirical human data (successfully). We dissected the model, post-hoc, to understand this difference between walking and running.
abstract; video

251
Shoulder mounted gyrsocopic shoulder prosthesis for assisting arm amputees during walking (8 minute talk)
251 Mitch Muller (Muller, Popovic)
We propose a compact arm prosthesis that uses a controlled moment gyroscope. It should restore some of the dynamics of the missing arm, in shoulder disarticulation and forequarter amputees, during walking and running. A benchtop prototype suggests that the actuator is well-suited to recreating the periodic moments exerted by the arm on the trunk. Before clinical trials, we need amounting scheme, to transmit load to the body, and to identify the best gait features to signal the control.
abstract; video

182
- (30 minute talk)
182 Max Donelan (Donelan)
I would like to be a team player and do what the organizers would most like me to do. However, I prefer to not present. I am most interested to learn about reinforcement learning, body-in-the-loop optimization, and fundamental principles of legged locomotion, be they energetic, mechanics or control.
video

316
Connecting Mechanical Properties to Biomechanical Performance of Prosthetic Feet via Fundamental Principles to Design Customized, Passive Devices (10 minute talk)
316 Amos Greene Winter V (Winter V, Olesnavage, Prost, Major)
We have a way to quantitatively connect the mechanical design of prosthetic feet to the biomechanical performance they induce. This way, we optimize passive, compliant feet to do physiological kinematics and kinetics. Our metric, called Lower Leg Trajectory Error (LLTE), compares motion of the prosthetic shank to able-bodied lower leg kinematics while the foot is loaded with physiological ground reaction forces. We'll explain the LLTE theory, our prototypes and our results. (See also the Prost talk)
abstract; video

228
Key Control Strategies Emerge in Spring Loaded Inverted Pendulum Traversal of Slippery Terrain (5 minute talk)
228 Samuel Pfrommer (Pfrommer)
Does numerical trajectory optimization of the SLIP (mass on spring) model on low-friction surfaces yield biologically meaningful control strategies? We formed a large, directly collocated optimization problem for hundreds of randomly generated scenarios. Some of the resultant trajectories matched control strategies previously discovered in slippery-terrain human-balance recovery.
abstract; video

320
Approximate explicit model predictive control for push recovery using mixed-integer convex optimization (12 minute talk)
320 Robin Deits (Deits, Marcucci, Manuelli, Koolen, Tedrake)
Explicit model-predictive control (MPC) is the solving of an entire family of optimization problems off-line and ahead of time. At run time the robot quickly looks up the pre-computed motions. So far, this approach has been limited to simple systems. We seek efficient approximations of the explicit MPC approach that is practical for humanoid robots. Making and breaking contact is a key problem.
abstract; video

356
An Experimental Comparison of Human-In-the-Loop Optimized Ankle Exoskeleton Control Strategies (10 minute talk)
356 Kirby Ann Witte (Witte, Collins)
For ankle exoskeletons during non-steady-state walking, which high-level control strategy results in the largest metabolic reductions? We plan human-in-the-loop optimization on three candidate controllers. These include proportional EMG, EMG with a muscle-tendon model, and time based control. We expect that, once optimized, EMG with a muscle-tendon model will be best.
abstract; video

215
Data-driven Geometric Gait Analysis (12 minute talk)
215 Ross L. Hatton (Hatton, Bittner, Ramasamy, Revzen)
Robots with many points of contact to the world have dynamics that are hard to model and optimize. We are combining data-driven Floquet analysis methods, for extracting system models from noisy data, with differential-geometric methods for evaluating optimality. Our goal is to make a gait-analysis and optimization paradigm that combines the experimental robustness of the data-driven tools with the theoretical rigor of the geometric tools.
abstract; video

294
Implementing Human-in-the-Loop Optimization on Prosthesis Emulators (2 minute talk)
294 Vincent Chiu (Chiu, Collins)
What is the best way to implement HitL optimization on a prosthesis emulator? We're still figuring it out, but here are a few challenges we've come across: parameter selection and the adaptation of the human subject.
abstract; video

319
Heuristic-based Adaptive Control of Ankle Exoskeleton Assistance Using Plantarflexor Electromyography (12 minute talk)
319 Rachel Jackson (Jackson, Collins)
People adjust their coordination patterns when walking with ankle exoskeletons, yet few exoskeleton control strategies take into account these human adaptations. Our new control technique optimizes the pattern of exoskeleton torque online using measured soleus muscle activity. The goal is to drive soleus muscle activity down towards zero and enabling co-adaptation between the device and the human. Experiments show that this new control approach can greatly reduce soleus muscle activity and, consequently, whole-body metabolic rate.
abstract; video

272
Robust walking with a simple IP model (8 minute talk)
272 Petr Zaytsev (Petr Zaytsev, Andy Ruina)
We aim to better understand the limits of robustness in walking to help the design of robust controllers. We use viability-theory concepts and a simple planar model, to assess all feasible states and control actions of the robot. We find a walking controller for Cornell Ranger that is, in some way, maximally robust.
abstract; video

394
An adaptive neural network learns to be a state estimator and central pattern generator (CPG) (8 minute talk)
394 Hansol Ryu (Ryu, Kuo)
We propose that animal nerve systems perform state estimation. They use the estimation to produce a rhythmic pattern. But it's not known whether and how a neural system can do this. A neural network was trained to be both a state estimator and a pattern generator. Such a network also adapts to changed conditions.
abstract; video

278
Locomotor sub-functions for design and control of locomotion (3 minute talk)
278 Maziar Ahmad Sharbafi (Sharbafi, Seyfarth)
We introduce the concept of locomotor sub-functions: Stance, Swing and Balance. This concept is used to help understanding of legged locomotion, including design and control of artificial locomotors. We have various implementations of this approach to explain human gait data, models, robots and even an exoskeleton.
abstract; video

303
Human running controller derived from steady-state running variability. (12 minute talk)
303 Nidhi Seethapathi (Seethapathi, Srinivasan)
How do people run without falling down in the presence of naturally-ocurring perturbations? We collected and mined nominally constant-speed running motion and running force data. We implemented the inferred controller on a simple point mass telescoping leg model. We uncovered the strategies used to modulate leg force and foot placement and found that these strategies stabilize the mathematical model for a large range of perturbations.
abstract; video

334
Mina V2 Powered Exoskeleton (7 minute talk)
334 Tyson Cobb (Neuhaus, Gines, Cobb, Craig, Smith, Griffin)
We built an exoskeleton with a powered ankle to compete in the 2016 cybathlon. The use of a powered ankle allowed us to walk faster, take longer steps, and decrease user fatigue. We attribute much of our success at the Cybathlon to the implementation of this design.
abstract; video

331
Humans falling in holes: joint-level mechanical response to a sudden perturbation in substrate height during vertical hopping. (10 minute talk)
331 Gregory S. Sawicki (Dick, Punith, Sawicki)
Our goal is to learn from human behavior how to make assistive devices work better in unsteady conditions. We measured human lower-limb response -- from joints to muscles -- as they recover from falling into a hole. Energy dissipation moves from the proximal to distal limb and the Achilles tendon protects the calf muscles from rapid stretch.
abstract; video

252
A Reduced Gait Model for Motion Prediction in the Clinic (4 minute talk)
252 Matthew Millard (Matthew Millard, Manish Sreenivasa, Martin Felis, Katja Mombaur, and Sebastian I. Wolf)
We are working to make model-based optimization a useful tool for a clinical orthopaedic team that treats children with gait disorders. We have made a reduced musculoskeletal model that makes it easy to identify and fit the model to the patient. We can predict the motions and forces of a healthy 10-year old boy walking. But, the foot-ground contact model needs improvement.
abstract; video

223
Shape and Size Optimization of a Passive Prosthetic Foot to Replicate Physiological Lower Leg Trajectory (5 minute talk)
223 Kathryn Michelle Olesnavage (Olesnavage, Winter)
I use compliant mechanism shape and size optimization to design a single-part prosthetic foot that, when acted upon by typical ground reaction forces, best replicates able-bodied kinematics for the lower leg.
abstract; video

239
Is a passive perturbation device assisting medial-lateral balance during walking? (8 minute talk)
239 Jean-Paul Martin (Jean-Paul Martin and Dr. Qingguo Li)
Can we provide horizontal stability assistance using a completely onboard system, a weight in a backpack that oscillates sideways? We had subjects walk with a a related device, an inverted pendulum, while collecting kinematic and kinetic data. We seek a relationship between the mass motion and a subject's leg work and energy changes, during the step-to-step transitions.
abstract; video

259
Journey of a Dynamic Walker (0.5 minute talk)
259 Keith William Buffinton (Buffinton)
I am interested in appropriately-complex dynamic models of walking robots, including compliances, yet simple enough for analysis. Addressing the related modeling questions builds on things I know about: flexible rotots using Kane's Method, vibration suppression, and nonlinear dynamics applied to gait.
abstract; video

179
Energy-robustness trade-offs in locomotion (30 minute talk)
179 Manoj Srinivasan (Srinivasan)
Are there any trade-offs between energy optimality and robustness/stability in locomotion and other simpler mechanical tasks? We use numerical optimization of a simple models to explore such trade-offs. We find that there are trade-offs --- and depending on the costs and uncertainties, robustness is improved by changing the nominal motion/posture.
video

236
The Energetic Effect of a Flexible Spine in Quadrupedal Robots (10 minute talk)
236 Yevgeniy Yesilevskiy (Yesilevskiy, Yang, Remy)
We explore whether a quadrupedal robot can receive an energetic benefit by adding an articulated powered spine. Using optimization, we compare the optimal motion of rigid and articulated models across a variety of speeds and gaits. The articulated model uses much less energy at high speeds, particularly when galloping.
abstract; video

376
An overview of bicycle balance rider control (25 minute talk)
376 Arend Schwab (Schwab)
How does a rider balance a bicycle? What sensors does he use to determine the state of the system, and which control actions does he use? We have some experiments that address these questions. For example, we found that riders detect lean, in part, by the feel of their handlebars.
abstract; video

301
New Hydraulic Humanoid Robot High Level Design (30 minute talk)
301 Jerry Pratt (Pratt, Ragusila)
Will present the high level design of a planned hydraulic humanoid robot. We want to approach the mass, volume, range of motion, and power of a human gymnast.
abstract; video

186
Linking mechanics and control (8 minute talk)
186 Oezge Drama (Heim, Grimmiger, Drama, Sprowitz)
How do bio-mechanical observations of legged animals explain the components of legged robots. How do legged-robot design choices map to motion primitives and control policies. We will implement different leg designs on hardware and in simulation and assessthe design's effect on learning and controller design.
video

340
Continuous Energy Optimization Controls Preferred Step Width in Human Walking (10 minute talk)
340 Sabrina Abram (Abram, Selinger, Donelan)
How much does the nervous systems of able-bodied people weigh energy-optimization when choosing step width. Our device uses closed-loop control to apply energetic penalties, as a function of step width, to shift the energy-optimal step width wider than that initially preferred. Subjects adapted their preferred step width to the new energy optimum.
abstract; video

324
Predicting Energy Cost using Portable Physiological Sensors (10 minute talk)
324 Kimberly A Ingraham (Ingraham, Ferris, Remy)
The benefits of an assistive device, like a prosthesis or an exoskeleton, are often measured by reduced energy use. But, measuring energy use in real time is hard, needing a bulky mouth and nose mask. I want to predict energy cost in real time using unobtrusive, portable, wearable sensors, like heart rate monitors and accelerometers. My work indicates that combinations of physiological signals are able to predict energy cost across subjects, especially if using a sliding-window average.
abstract; video

289
Human-in-the-loop Optimization of Exoskeleton Assistance (30 minute talk)
289 Steve Collins (Collins)
It is hard to design exoskeletons and prostheses that actually help users. We develop algorithms that automatically tune device behavior to maximize human performance. This substantially improves the energy cost of assisted walking and running compared to hand-tuning.
abstract; video

293
High-Order Contact-Implicit Trajectory Optimization (10 minute talk)
293 Zac Manchester (Manchester, Kuindersma)
Trajectory optimization for walking systems shouldn't require contact sequences to be pre-specified by the user. We made a new family of contact-implicit trajectory optimization methods. These use ideas from discrete mechanics that meet this goal while also achieving high-order integration accuracy. Examples will be given from a third-order accurate method that is competitive with hybrid direct collocation schemes.
abstract; video

225
Neural Control Versus Mechanically Intrinsic Control of Powered Ankle Exoskeletons (8 minute talk)
225 Jeffrey Russell Koller (Koller, Remy, Ferris)
Two ways to control lower extremity assistive devices: 1) driving the control with mechanically-intrinsic measurements, and 2) driving the control with neural measurements. To date, there has not been a fair comparison of the two. So we systematically compared them with bilateral ankle exoskeletons. We tentatively found that using neural measurements to drive the control might better benefit rehabilitation, and using mechanically intrinsic measurements might better benefit metabolic reductions.
abstract; video

234
Linking Mechanics and Learning (7 minute talk)
234 Steve Heim (Heim, Grimmiger, Drama, Sprowitz)
How should natural dynamics (i.e. mechanics) be designed to make learning control easier/more effective? We want to systematically test different designs (i.e. compliance, leg segmentation etc.) both in hardware and software and specifically see how this affects convergence for Reinforcement Learning (RL). We hope to identify effective mechanical design "blueprints" that are particularly amenable to RL.
abstract; video

329
QWOP-timization II: Can we make a control policy in spite of unlimited examples? (10 minute talk)
329 Matthew Sheen (Sheen)
QWOP (four letters on a keyboard) is a popular video game, a "running simulator" of sorts. QWOP is hard to play. I want to make a computer-controlled physical robot which sits at a computer and plays the game. So far, I can make lots of open-loop trajectories. Now I'm working on closing the loop and making a control policy, a process made difficult by many unobservable states, an inconsistent game clock, and other issues.
abstract; video

275
Towards measuring agility for legged, terrestrial locomotion (5 minute talk)
275 Peter Eckert (Eckert, Ijspeert)
What is agility and how can we measure it? Inspired by dogs and other animals, we developed a benchmark. The benchmark is easy to implement and allows comparison between robots. It's not out yet. I really hope that people will use our benchmark to compare robots more clearly and thus more-efficiently drive robot development.
abstract; video

308
Predicting steady and transient locomotor dynamics based on the principle minimizing energy cost (15 minute talk)
308 Monica A Daley (Daley, Hubkicki)
Do animals minimize energy in transient behaviors, like accelerating, turning, and going around obstacles? Our simple biped model, using trajectory optimization to minimize cost of transport, predicts steady and transient motionss. Perhaps cost-minimizing will explain animal motion without the need to invoke factors like force constraints and stability.
abstract; video

180
Gyroscopic balance assistance: First results and remaining challenges (10 minute talk)
180 Andy Berry (Berry, Lemus, Vallery)
Recovery from neuromotor disorders often involves extensive gait training, often with robotic aids. Few such aids help with balance. We have designed and built two wearable backpack-like devices that use gyroscopic actuators which assist balance during standing and locomotion. We are designing controllers and testing the devices. I'll showpreliminary results and ask the audience for advice about control strategies.
video

309
Deriving the human walking controller from perturbed walking data (15 minute talk)
309 Varun Joshi (Joshi, Srinivasan)
We would like to know the controller humans use to stabilize themselves while walking. To do this, we asked subjects to walk on a treadmill and measured how their kinematics and ground reaction forces changed in response to large torso perturbations. Using this data, we derived controllers for two different models of human walking.
abstract; video

306
Predictive Planning Based on Reactive Control (12 minute talk)
306 Patrick James Clary (Clary, Hurst)
Cassie, our newest bipedal, walks using controllers that react to disturbances, instead of preemptively planning around obstacles. These controllers can not execute planned trajectories. But experiments and simulations show that our reactive controllers yield predictable-enough behavior for use as in model-predictive planning. We will develop thismovement-control approach more so robots like Cassie can go quickly in complex environments, while retaining the robust disturbance recovery of the underlying reactive controller.
abstract; video

198
Can we make the statement "To balance keep your support under your center of mass" precise? (10 minute talk)
198 Andy Ruina (Ruina (maybe also Vallery))
I usually say that balance of bicycles, walkers and runners is accomplished by moving the center of mass (CoM), and/or the support point, so that the two are vertically aligned. But there are situations that one could call 'balanced' where this doesn't hold. How precise is the need for the support point to be under the CoM?
video

321
Robots Should Have Brakes, That Work Like Muscle (20 minute talk)
321 Chris Atkeson (Christopher G. Atkeson)
Recent work in muscle mechanochemistry suggests that muscle behaves very differently mechanically when lengthening vs. when shortening. I am exploring the implications for robot control: cheap robots should have physical brakes.
abstract; video