Humanoid Robot Soccer 101 Thomas Rfer Cyber-Physical Systems German - - PowerPoint PPT Presentation

Thomas Röfer Cyber-Physical Systems German Research Center for Artificial Intelligence (DFKI)

Humanoid Robot Soccer 101

1

RoboCup 2013: SPL Semifinal

2

RoboCup 2013: Statistics

30 11 9 7 13 1 22 15 9 1 7 2 7 8 2 11 4 22 62 22 5 2 4 1 6 4 8 11 9 8 1 40 15 5 3 12 1 4 10 1 28 7 9 5 12 26 18 3 90 6 9 7 8 3 22 11 17 25 34 11 4 2 1 8 16 3 27 12 12 7 13 2 5 11 1 6 5 14 1 4 8 10 1 15 1 7 4 16 14 4 15 16 7 18 8 1 8 11 2 5 3 15 26 26 9 5 13 8 32 5 12 2 2 5 14 15 11 12 14 1 4 3 9 6 12 8 7 8 13 18 10 17 11 12 1 10 2 27 6 25 18 22 1 16 19 5 3 54 12 1 11 5 1 1 14 1 1 7 17 2 10 2 6 1 6 3 7 7 16 4 16 27 21 4 17 10 2 7 39 36 12 8 Goals Out by Ball Holding Fallen Robot Illegal Defender Inactive Player Leaving the Field Request for PickUp Player Pushing Playing with Hands Austin Villa Austrian Kangaroos B-Human Bembelbots Berlin United Cerberus DAInamite Dutch Nao Team Edinferno Kouretes Mi-Pal Nao Devils Dortmund Nao-Team HTWK Northern Bites NTU RoboPAL RoboCanes RoboEireann rUNSWift SPQR Team TJArk UChile Robotics Team UPennalizers

3

Standard Platform League

• Aldebaran Robotics NAO
• 21-25 degrees of freedom
• Height 57cm, weight 5 kg
• Different sensors, on-board PC (1.6GHz Atom)
• Soccer Competition
• 5 vs. 5
• Robots are fully autonomous
• Field size 9 m x 6 m

4

60 Hz 100 Hz

s

Controlling a Soccer Robot Cognition Motion

• Perception: What do I see now?
• World Modeling
• Where am I?
• Where are objects currently not perceived?
• What speeds do objects have?
• Behavior Control: What to do?
• Sensing: What am I feeling?
• Motion Control: Walking, kicking, standing up, looking

5

Perception

6

Perception: Grid-based Scanning and Specialists

7

Perception: Determining Distance

distance from size distance from bearing

8

Sensing

• Center of Mass
• Ground contact
• Falls (with direction)
• Robot is falling
• Torso pose
• Camera poses

9

Sensing: Torso Pose

• Unscented Kalman Filter
• Forward kinematics
• Calibrated gyroscopes
• Compensation for gyroscope’s bias drift

10

Sensing: Calibrating Camera Pose

• Before calibration
• Misplaced camera
• Backlash in joints
• After calibration
• Camera roll / tilt
• Overall body roll

11

Perception and Sensing: Synchronization

• Rolling shutter (CMOS technology exposes pixel-by-pixel)
• Time differences between images and joint angles
• Correction
• Using head joint 


velocities

• Only perceptions, not 


whole image

Image taken by Bioloid robot

12

World Modeling: Self-Localization, Ball

• Self-localization
• Particle filter with 16 


Unscented Kalman Filters

• Side confidence and own side model
• Use ball for disambiguation
• Ball modeling
• 6 Kalman Filters for static ball
• 6 Kalman Filters for rolling ball

13

World Modeling: Obstacles

• Sonar-based
• Overlapping measurement areas
• 2-D evidence grid of

measurement history

• Vision-based
• Edges between field 


and robots

• Obstacle wheel

14

Motion Control

• Walking
• Kicking
• “Special actions”
• Getting up
• Head control
• Scan interesting points on the field
• Hard-coded modes

15

Motion Control: Walking

16

• Omni-directional
• Modeling single support phase 


as linear inverted pendulum

• Balancing with difference between observed and planned COM

Motion Control: Walking and Kicking

92°/s 31cm/s 22cm/s 12cm/s

17

Motion Control: Balanced Dynamic Kicks

18

• Modeled as a sequence of Bezier curves
• 2x foot positions, 2x foot rotations, 


2x arm positions

• Transitions continuous in place and

gradient

• Control points are adapted during kick
• Balancing based on
• Preview of COM
• Gyroscopes

Motion Control: Balanced Dynamic Kicks

19

Behavior Control: Hierarchical State Machines (Options)

20

Behavior Control: States and Decision Trees

21

Behavior Control: CABSL – C-based Agent Behavior Specification Language

• Directly compiled by C++ compiler
• Modeling behavior with hierarchical state

machines (options)

• Each option contains states
• Each state contains
• conditional transitions to other states
• actions (C++, calls to other options)
• Each option can only switch its state once

per execution cycle

22

• ption(goaliePlaying) {

initial_state(stayInGoal) // ... state(getToBall) { transition { if(ball.notSeenFor > 500 || ball.distance > 600) goto returnToGoal; else if(ball.distance < 150) goto clearBall; } action { GoToBall(); } } }

Behavior Control: CABSL – Special States and Symbols

• initial_state (mandatory): Option returns to this state when it

was not executed in the previous cycle

• target_state: Caller's symbol action_done becomes true if

the last sub option it called reaches this state

• aborted_state: Caller's symbol action_aborted becomes

true if the last sub option it called reaches this state

• option_time: How long since entering the initial_state?
• state_time: How long since entering the current state?

23

Behavior Control: Team Play

• Roles: Striker, supporter, breaking supporter, defender, keeper
• Global world model
• Global ball for role switching
• Teammate positions for path 


planning

• Joint actions
• Kick-off, passing
• Synchronized ball tracking and searching

24

Behavior Control: B-Human 2013

25

Conclusions

• Doing the right things
• Grid-based vision
• Probabilistic world modeling (often based on textbook methods)
• Hierarchical state machines for behavior control
• Balanced walks and kicks
• Doing things right
• Keeping 60Hz/100Hz
• Synchronization and calibration

26

E x e r c i s e : w w w . t z i . d e / s p l / b i n / v i e w / W e b s i t e / H S S 2 1 3