Push-Recovery ry Stability of Biped Locomotion Yoonsang Lee 1 Kyungho Lee 2 Soon-Sun Kwon 3 Jiwon Jeong 1 Carol O’Sullivan 4 Moon Seok Park 5 Jehee Lee 2 1 Samsung Electronics Co., Ltd. 2 Seoul National University 3 Ajou University 4 Disney Research 5 Seoul National University Bundang Hospital
Biped Controller Design Two fundamental goals: • Human-likeness (i.e., normal gait) • Robustness against perturbation [Lee et al. 2010]
Normal Gait • Gait pattern that humans normally adopt • More energy efficient than other gaits • Many questions still remain regarding stability
Questions | | Gait it St Stabil ilit ity • Under what conditions is human gait more stable? • Is normal gait more stable than abnormal gait?
Questions | | Bip iped Control • How much are existing biped controllers human- like? • Are they as stable as human locomotion? • Also influenced by the gait factors that affect humans?
Goal Correlation between gait and stability • which aspects of gaits affect balance-recovering capabilities • how these factors affect to computer-simulated bipeds Better understanding of existing biped controllers & how to improve controller design
Background | Measurin ing St Stabil ilit ity • Standing postures stability [Rogers et al. 2001] [Brauer et al. 2001] • Gait stability [Bauby and Kuo 2000] [England and Granata 2007] [Bruijn et al. 2009] • Our work: Recovery response for moderate push while walking
Background | | Bip Biped Co Control • Bipedal Locomotion [Wang et al. 2010, 2012] [Geijtenbeek et al.2013] [Mordatch et al.2013] [Kwon and Hodgins 2010] [Coros et al. 2010] [de Lasa et al. 2010] [Yin et al. 2007] [da Silva et al. 2008] … [Wang et al. 2010] • Stability to external pushes [Lee et al. 2014] [Wang et al. 2010] [Lee et al. 2010] [Yin et al. 2007] • Stomping gait controllers were better - affect of gait type?
Design of f Experiments Top view Side view • Stability measure : Lateral deviation from a line • Experiments with human participants & simulated biped
Experiments wit ith Human Part rticipants • 30 healthy participants ( 15 males / 15 females) • Measure 31 variables: • Level of crouch (0 ˚,20˚, 30 ˚, 60˚) • Stride length • Walking speed • Timing, magnitude and direction of the push • Height, weight, BMI, and leg lengths • Angle of the feet at ground contact • …
Collected Data fr from Experiments Group 1 Group 2 • The lateral displacement • The lateral displacement peaked within one step peaked within three steps • Easily recovered balance • Experienced difficulties recovering balance
Collected Data fr from Experiments • Multivariate, repeated, unbalanced, and clustered • Outcome is a time series of complex articulated structures • Measured twice on the same outcome • Difficult to collect well-balanced data with so many variables • Linear Mixed Model (LMM) • General method for handling the between- and within- subject variability in the repeated, unbalanced data. • Model selection • LMM with walking speed, push magnitude and timing minimize AIC & BIC
Result | | Overall ll Analy lysis is • The gait detours less if (for both human & simulation) • Level of crouch It crouches mild • Walking speed Walks faster • Magnitude of push Push is weaker • Timing of push Push happens later • Height, weight, BMI were irrelevant.
Simulation Experiments | Bip iped Con ontroll ller • Data-driven biped controller [Lee et al 2010] • Imitating reference gait
Simulation Experiments | Bip iped Con ontroll ller • Feedback rules to maintain balance • stance hip & ankle, swing hip & foot height • Feedback parameter vector θ Without feedback ( θ = 0) With hand-tuned parameter θ
Simulation Experiments | | Par arameter Optimization • In the original work, θ was manually tuned for each motion • In this work, θ is optimized to handle gait variations penalizes early falling penalizes excessive parameters favors better tracking • Evaluated with simulation of normal gait & 60 ° crouch gait • 10s duration, with push in the middle • Enough to deal with a wide variety of gait variations reliably
Simulation Experiments | | Push sh Experiments • Gait variations with random variables • Level of crouch, stride length, walking speed, the magnitude and timing of push • Multivariate normal distributions • Using motion displacement mapping and timewarping • Fed into the optimized controller as reference motion • Simulated biped is pushed while walking
Result | | Le Level l of f Cr Crouch Group 1 Group 2
Result | | Energy Exp xpendit iture of Sim Simula lated Walk lkin ing • Cost of Transport : required energy to move unit distance • Humans : 1.05m/s ~ 1.4 m/s • Normal gait simulation: 1 m/s • Crouch 30 simulation : 0.76 m/s
Result | | St Stri ride Length & St Stri ride Frequency • Additional analysis - the gait detours less if • Stride Length Shorter stride length • Stride Frequency Higher frequency • Opposite to the results of previous studies • [Bauby and Kuo 2000; Bruijn et al. 2009] [England and Granata 2007] • Reason - Different stability estimates • Prevs ’ - Kinematic variation (Lyapunov exponents) • Ours - Resilience against impulsive perturbation
Discussion | | Su Summary ry • Significant factors affecting push -recovery response • Level of crouch, walking speed, push force and push timing • Simulated biped is affect by same factors of human • Quantitatively, human participants performed better than simulation.
Discussion | Im Impli licatio ion • How much are existing biped controllers human- like? • Existing biped controllers behave quite similar to humans qualitatively
Discussion | Im Impli licatio ion • About evaluating robustness of biped controllers • Normalization for push-recovery test of biped controllers
Acknowledgements • Thanks to • All members of SNU Movement Research Laboratory • Seung Yeol Lee and Mi Seon Yoo from Seoul National University Bundang Hospital • Funding • National Research Foundation of Korea (NRF) No.2011- 0018340 , No. 2007-0056094.
Push-Recovery Stability of f Bip iped Lo Locomotion Yoonsang Lee 1 Kyungho Lee 2 Soon-Sun Kwon 3 Jiwon Jeong 1 Carol O’Sullivan 4 Moon Seok Park 5 Jehee Lee 2 1 Samsung Electronics Co., Ltd. 2 Seoul National University 3 Ajou University 4 Disney Research 5 Seoul National University Bundang Hospital
Recommend
More recommend
