Towards Robots that Move and Interact Like Humans

1
Towards Robots that Move and Interact Like Humans

• Katsu Yamane

Carnegie Mellon University
Disney Research, Pittsburgh
2
About Disney Research, Pittsburgh
3
Disney Research

• http//www.disneyresearch.com/
• informal association of research labs in The Walt
  Disney Company
• Disney Interactive Media Group
• Disney Research, Pittsburgh (DRP)
• Disney Research, Zurich (DRZ)
• Pixar Animation Studios
• Walt Disney Animation Studios (WDAS)
• Walt Disney Imagineering (WDI)

4
Disney Research, Pittsburgh (DRP)

• people
• director Jessica Hodgins computer graphics,
  robotics
• senior research scientists
• Iain Matthews computer vision, face tracking and
  animation
• Ivan Poupyrev (Aug. 2009) human-computer
  interaction
• Katsu Yamane humanoid robots, animation,
  biomechanics
• postdocs
• Josh Griffin, Raffay Hamid, Michael Mistry,
  Takaaki Shiratori, Edilson de Aguiar
• 1 administrative coordinator
• 3 visiting professors / 9 interns

5
Disney Research, Pittsburgh (DRP)

• research areas
• robotics
• humanoids
• human-robot interaction
• animation
• motion capture
• facial animation
• human perception of animated motion
• interfaces
• audio, vision, novel sensors
• data mining, visualization

6
Robotics Research at Disney

• (probably) largest number of humanoid robots
  directly used in business everyday
• many real-world problems
• many potential innovations
• human shape and natural motion are required
• various test environments
• indoor, controlled
• outdoor, unstructured

7
Research Overview
8
My Background

• 1997-2002 PhD student, University of Tokyo
• advisor Prof. Yoshihiko Nakamura
• simulation and motion synthesis of humanoid
  robots
• musculoskeletal human models
• March 2002 PhD in Mechanical Engineering
• 2002-2003 Postdoctoral Fellow, Carnegie Mellon
  University
• with Prof. Jessica Hodgins
• control of motorized marionette using motion
  capture data
• motion capture database and planning for
  character animation
• 2003-2008 Assistant/Associate Professor, Univ.
  of Tokyo
• human neuro-musculo-skeletal model and dynamics
• motion database
• Oct 2008- Disney Research, Pittsburgh
• humanoid robot control from motion capture

9
Main Research Interest

• How do we coordinate our movement?
• issues
• modeling (body/control/motion)
• algorithms for motion analysis and simulation
• measurement (mocap, force, EMG, fMRI)
• parameter identification
• applications
• robotics
• graphics
• sport science
• medical science

10
Talk Overview
11
Talk Overview
12
Humanoid Simulator

• two projects with AIST (1999-2001, 2005-2008) to
  develop a simulation platform for humanoid robots
• requirements
• articulated rigid bodies
• collision/contact model
• general polygonal objects
• concerns
• precision
• numerical stability
• computational efficiency
• now open source!

OpenHRP3 http//www.openrtp.jp/openhrp3/en/index.h
tml
13
Simulation of Humanoid Robots

• structure-varying kinematic chains
• the link connectivity may change during the
  motion
• need to handle general closed kinematic chains

Nakamura, Yamane IEEE TRA 2000
14
Scalablility

• decrease the complexity
• O(N2)?O(N) (N number of joints) Featherstone
  1987 Rothenthal 1990 Baraff 1996
• employ parallel processing
• O(logN) on O(N) processors Fijany et al. 1995
  Featherstone 1999 Anderson, Duan 2000
• Assembly-Disassembly Algorithm (ADA)
• Yamane, Nakamura ICRA 2001, 2002
• applicable to any open/closed kinematic chains
• automatic scheduling for parallel processing
  Yamane and Nakamura, RSS 2007
• final results are equivalent to Featherstones
  algorithm Yamane and Nakamura, IJRR 2009

15
Application to Human Models
Dual Pentium Xeon 3.8GHz x 2
16
Collision/Contact Models

• spring-damper (penalty-based) models
• easy to implement
• - difficult to find appropriate parameters
• - small timestep for integration (numerically
  unstable)
• rigid-body (constraint-based) models Stewart
  2000
• relatively precise (in theory)
• numerically stable (in theory)
• - difficult to implement very few existing
  libraries work for humanoid robots sensitive to
  numerical errors
• impulse-based models Mirtich 1995
• numerically stable
• - not good for continuous contact (e.g. stacking)

17
Our Approach Yamane, Nakamura RSS 2008

• develop a robust pivot-based solver for linear
  complementarity problems (LCP)
• improve Lemkes algorithm Murty 1988
• maintain all pivot candidates
• search for the optimal pivot sequence using
  dynamic programming
• collision/contact model by LCPADA
• extend the formulation of Stewart and Trinkle
  2000 to articulated rigid bodies

18
Linear Complementarity Problem (LCP)

• example normal direction

force
19
Experiments

• humanoid robot ut-m2 Magnum Sugihara 2005
• tap-dancing Yamamoto 2006

20
Experiments

• comparison with Lemkes algorithm

21
Experiments

• comparison with hardware

22
Experiments

• with closed loop (toe joint)

23
Non-Humanoid Examples
24
Talk Overview
25
UT-Poser Flexible IK

• develop a new tool for animators (with Sega)
• intuitive interface for manipulating characters
• interactively generate natural postures
• do not generate a sequence
• any number of constraints
• link position and/or orientation
• reference joint angles
• joint motion range
• technical difficulties
• singularity
• inconsistent constraints

Yamane, Nakamura TVCG 2003
26
UT-Poser Flexible IK

• solution
• singularity-robust inverse Nakamura and Hanafusa
  1986 Maciejewski 1990
• two priority levels hard/soft
• applications
• SEGAAnimaniumTM 2003 International 3D Awards
• motion capture

27
Talk Overview
28
Adapting Human Motion

• to deal with
• differences between human/robot bodies
• limb length, mass, actuation,
• new environment / constraints
• collision avoidance, task,

29
Issues

• different kinematics / dynamics
• may be physically infeasible
• may not be able to accomplish the task
• what can be modified / what has to be preserved
• joint trajectory, endeffector trajectory, contact
  force,
• usually task-dependent
• human adaptation
• acquired by practice
• once acquired, easily adapted to various scenes
• sophisticated motor control? memory?

30
Related Work

• common problem in robotics and animation
• usually formulated as an optimization problem
• how to solve the optimization problem?
• mathematical optimization Gleicher et al. 1997
  Ude et al. 2004 Liu et al. 2005
• learning Kuniyoshi et al. 1994 Atkeson, Schaal
  1997 Bentivegna et al. 2004
• task and skill Nakaoka et al. 2007

31
Talk Overview
32
Dynamics Filter

• convert a physically infeasible motion to a
  feasible one
• example

original
converted
Yamane, Nakamura TRA 2003
33
Dynamics Filter

• convert a physically infeasible motion to a
  feasible one
• reasons for infeasibility
• measurement error
• different kinematic/dynamic parameters
• manual editing
• technique
• obtain desired accelerations by a feedback
  controller in joint/Cartesian spaces
• project the accelerations onto feasible space by
  local (online) optimization
• limitation can only cope with small differences

34
Dynamics Filter Concept
acceleration
solution space of dynamics equation
contact forces
joint torques
desired acceleration (from controller)
locally optimized solution
unilateral contact forces considered
35
Dynamics Filter Examples
filtered
captured
hip trajectory edited
different mass property
36
Talk Overview
37
Synthesizing Manipulation Tasks

• adapt example motions to
• new objects
• new environment
• new task (start/goal positions)
• new character

Yamane, Kuffner, Hodgins SIGGRAPH 2004
38
Synthesizing Manipulation Tasks

• combination of model and data
• motion planning (RRT) LaValle and Kuffner 2000
  and inverse kinematics (UTPoser) Yamane and
  Nakamura 2003
• relatively small data set (four pick-and-place
  examples)
• limitations quasi-static, no stepping
• results

39
Synthesizing Manipulation Tasks

• using motion capture to bias IK

40
Talk Overview
41
Balance Control

• so far how to synthesize new reference motions
• in real robots
• disturbances
• model uncertainty
• maintain balance while tracking a reference
  motion
• often conflict with each other
• feedback control with sensors
• sensor noise

42
Controller Structure

• tracking and balancing during double support
• balancing keep center of mass around reference
  position
• tracking follow joint trajectory from motion
  capture

motion clip
reference joint trajectory
joint torque command
balance controller
tracking controller
simulator
robot
current joint angles
Yamane, Hodgins IROS 2009
43
Balance Controller

• linear quadratic regulator (LQR) designed for a
  simplified humanoid model

center of mass (COM)
center of pressure (COP)
inverted pendulum (IP) model
full humanoid model
reference COM

IP model
balance controller
desired COP
-

observer
-
estimated COM
measured COM/COP
44
Tracking Controller

• joint controller tracks motion capture
• optimization computes optimal joint torque
• quadratic cost function with analytical solution

minimize (joint acc error)2 (COP error)2
(foot acc error) 2 subject to complete robot
dynamics
reference joint trajectory
joint controller
optimization
joint torque command
foot controller
desired COP (from balance controller)
current joint angles/velocities
45
Results Im a little teapot
46
Features

• general
• same model/controller for different reference
  motions
• online
• simple data cleanup for flat contact of
  supporting foot (feet)
• analytical solution for optimization
• low gain
• for exact model
• for wrong mass parameter
  example
• adapt to disturbances

47
Another Example
48
Talk Overview
49
Human Motion Control

• a key for interactive robots
• hierarchy of controllers
• high-level motor control in the brain
• large (100ms) delay too slow to cope with
  disturbance?
• low-level reflex in the spinal cord
• sensors
• somatosensory information (muscle length,
  tension)
• touch, temperature, pain
• smaller (30ms) delay
• most humanoid controllers run at or faster than
  1KHz
• some sophisticated mechanism?

50
Towards Human Motion Control

• models for analysis
• detailed musculoskeletal model for estimating the
  somatosensory information Yamane et al. ICRA
  2005
• neuromuscular network model with somatosensory
  reflex Murai et al. EMBC 2008

51
Musculoskeletal Model

• skeleton 155 DOF
• 200 bones ? 53 groups
• composed of mechanical joints
• hand/foot fingers not included
• actuator more than 1,000 wires
• 997 muscles linear actuators
• 50 tendons connect muscles and bones
• 117 ligaments constrain the bones
• algorithms
• inverse kinematics ? muscle length / velocity
• inverse dynamics ? muscle tension estimation

52
Musculo-Tendon Network

• mass-less, zero-radius wires with via-points

53
Actuation

• mapping wire tensions to joint torques
• moment arm limited to 1DOF joints, planar
  motions
• generalization by Principle of Virtual Work
  dAlemberts Principle

wire tensions
54
Estimating Wire Tensions

• numerical optimization to resolve redundancy
• consider EMG data for physiological reality
• linear programming formulation

minimize
subject to
error of the mapping equation
error from a reference tension
? minimizes total force
compute from EMG data
muscles can only pull
55
Examples
toe walk
heel walk
56
vertical contact force left right
left vastus lateralis left achilles
toe walk
heel walk
57
Neuromuscular Network Model

• inputs and outputs
• inputs motor command signals at spinal nerve
  rami
• outputs muscle tensions
• two paths
• CNS?PNS descending pathway
• PNS?PNS ascending and descending pathways
• (somatic reflex network)

peripheral nervous system (PNS)
central nervous system (CNS)
58
Identifying the Motor Command Signals

• independent component analysis (ICA)
• estimate mutually independent signal sources
• order of the independent signals is undefined
• dimension of ?
• 120, the number of relevant
• spinal nerve rami, is enough!

muscle tensions
independent signals
motor command signals
59
Neuromuscular Network Model

60
Neuromuscular Network Model
120 independent signals
61
Neuromuscular Network Model
all-to-all connection
anatomical connection
62
Neuromuscular Network Model
anatomical connection
63
Neuromuscular Network Model
muscle tensions
64
Neuromuscular Network Model
muscle length / force sensors
65
Neuromuscular Network Model
anatomical connection with time delay (30ms)
66
Identification

• training data
• walk (2000 frames, 10 seconds)
• muscle tensions from inverse dynamics
• independent signals from ICA
• training
• 5000 cycles
• error
• average 2.59
• variance 0.34

67
Results Weight Parameter of Reflex Loop

• from Iliacus agonist for hip flexion

classified as agonist muscles for hip flexion in
sports science
68
Results Weight Parameter of Reflex Loop

• from Tensor Fasciae Latae agonist for hip flexion

classified as antagonist muscles for hip flexion
in sports science
69
Results Patellar Tendon Reflex

• patellar tendon reflex stretch reflex of
  quadriceps

hit!
70
Summary

• efficient numerical tools
• dynamics simulation of humanoid robots
• UT-Poser flexible inverse kinematics
• adapting human motion
• dynamics filter adaptation to small differences
  in kinematics and dynamics
• synthesizing manipulation tasks adaptation to
  different kinematics and environment
• balancing and tracking adaptation to
  disturbances, uncertainty
• understanding human motion control
• musculoskeletal and neuromuscular network models

71
Acknowledgements

• University of Tokyo
• Yoshihiko Nakamura (Dept. of Mechano-Informatics)
• Tomotaka Yamamoto (School of Medicine)
• Akihiko Murai (PhD student)
• Masaya Hirashima (postdoc)
• funding JSPS, NEDO, IPA
• Carnegie Mellon University
• Jessica Hodgins
• James Kuffner
• Ben Brown
• Graphics Lab
• funding NSF

72
Questions?