Template Composition for Synthesis of New Behaviors from Simpler Constituents

This talk was given at the IROS 2018 templates workshop https://iros18wsdll.netlify.app/talks/

Transcript

1. Template Composition for Synthesis of
   New Behaviors from Simpler Constituents
   Avik De
   Post-doctoral scholar; Ph.D. with Daniel Koditschek (2017), University of Pennsylvania
   Cofounder & CTO, Ghost Robotics (Philadelphia)
   ARL/GDRS RCTA W911NF-1020016
   AFRL grant FA865015D1845 (subcontract 669737-1)
   ONR/OSD Vanevar Bush Fellowship N00014-16-1-2817
   

   View Slide

2. Legged robotics in the field: status update
   Vision-60 (30kg, 12dof)
   Research Scientific Commercial
   Minitaur (6kg, 8dof direct drive)
   Jerboa (3kg, 4dof)
   

   View Slide

3. What do we want from our control algorithms? (long-term view)
   [Raibert (1980’s)]
   [DRC (2017)]
   Robust against
   • State perturbations
   • Feedback control, MPC, …
   • Parameter perturbations
   • Modeling errors
   • “robust” control, model reduction
   Verifiable
   • Repeatability, accountability
   • Convergence
   • Local/global stability
   Generalizable solns
   • Across tasks
   • Across platforms
   • “online” solution, …
   …Modularity
   [Seyfarth (2015) [Schaal (2006)]
   

   View Slide

4. Inspiration from Biology
   [Ting et. al. (2016)]
   Biology Robotics
   Motor abundance Multiple solutions
   Preferred patterns of
   coordination are
   modulated across
   movements
   Reuse modules across
   behaviors
   Facilitate multifunctionality
   (same muscle, different
   function)
   Manage complexity of
   high-dimensional robots
   [Ting (2007)]
   

   View Slide

5. • Gait control synthesis using compositions of templates
   • Jerboa [ICRA 2015] [ISRR 2015] [IROS 2016] [IROS 2018]
   • Minitaur [IJRR 2018]
   • Role of design in enabling compositional control
   • Direct-drive machines [RAL 2016]
   • Multifunctional design [in prep]
   • Analysis for understanding when it works, and providing guarantees
   • Hybrid averaging [IJRR 2018]
   • Symmetry paper [in prep]
   Outline = Research Agenda
   

   View Slide

6. Gait control synthesis using
   compositions of templates
   

   View Slide

7. Applications
   [Shamsah, De, Kod (IROS 2018)]
   [De, Kod in IJRR (2018)]
   Minitaur bounding, trotting, pacing, pronking Jerboa tailed hopping
   [De, Kod (ICRA 2014)]
   

   View Slide

8. Control as reduction (anchoring) + composition
   Classical (analytical) view
   Templates are themselves compositions
   [Full and Koditschek (1999)] “Templates and anchors…”
   

   View Slide

9. Reduction (anchoring) exposes templates: reactive walking
   

   View Slide

10. Reduction (anchoring) exposes templates: hopping, running
    

    View Slide

11. Designing robots for
    compositional control
    

    View Slide

12. Role of design in enabling simpler control strategies
    Design for composition?
    [Shamsah, De, Koditschek (IROS 2018)]
    Tuesday,
    October 2,
    2018
    18:00-
    18:50
    Session
    TuETS4
    Chair Yoshida,
    Eiichi
    (National Inst.
    of AIST), Co-
    chair Ramirez-
    Amaro,
    Karinne (Inst.
    for Cognitive
    Systems.
    Tech. Univ.
    München)
    Legged
    Robots II −
    TuETS4
    Room 2.L2
    Analytically-guided tailed hopper design—IROS 2018
    [Raibert (1980s)]
    

    View Slide

13. Hierarchical synthesis strategy
    [Shamsah et al (IROS 2018)] [Libby et al (TRO 2016)]
    [Libby et. al (Nature 2012)]
    

    View Slide

14. Preliminary observations about parameter regions
    Orthogonal parameter gradients
    • “Decoupled parameters”
    • Cross-product instantiation
    Feasible parameter regions
    • Composition must lie in intersection
    • Affects possible template behaviors
    [De, Kod (in prep)]
    Parameter space
    Contours of
    performance metric
    

    View Slide

15. Why does it work?
    

    View Slide

16. (Hybrid) dynamical averaging as a verification tool
    Hybrid averaging
    [De, Burden, Kod in IJRR (2018)]
    [Raibert (1986)]
    Q: When can you do this?
    [De (2017)]
    

    View Slide

17. • Not merely symmetry of
    trajectories; property of dynamics
    [Altendorfer et. al. (2004)]
    • “Reversible systems” ≈
    “Hamiltonian systems” [Lamb &
    Roberts (1998)]
    • Analytical benefit: find the
    “neutral” limit cycles
    Time-reversal symmetry
    [Raibert (1986)]
    or, why averaging is particularly suited to analysis of locomotion
    Neutral orbit
    Non-neutral
    [Raibert (1986)]
    

    View Slide

18. Putting together coupled systems as if they were decoupled
    Only terms with even* symmetry remain:
    Shank energy →
    Leg angle energy →
    Using all the prior ideas [De & Koditschek (in prep)]
    — Full dynamics
    — Averaged
    

    View Slide

19. Thanks for your time
    

    View Slide