Towards Segway Soccer

1
Towards Segway Soccer
Pickup Segway teams in adversarial environments

• Prof. Manuela Veloso, PI
• Dr. Brett Browning
• Dr. Paul E Rybski
• Thanks to
• Jeremy Searock, Michael Sokolsky, Betsy Ricker,
• David Rozner, Dinesh Govindaraju, Ling
  Xu
• Carnegie Mellon University

2
Pickup Segway Soccer

• Our goal
• Humans and robots competing with and against one
  another in soccer-like competitions with
  identical capabilities
• Teams pickup players dynamically
• No prior knowledge of teammate capabilities,
  skills, or strategies

Robots must model unknown teammates as well as
opponents in a concrete adversarial task
3
Modular, Accessible Hardware
USB Camera(s)

• Perception
• Color vision via USB Camera(s)
• Two moderate range laptops
• One for each of perception and action
• Communication via wired Ethernet
• Choice of where to run cognition
• Actuators
• Segway RMP platform
• Pneumatic kicker for ball manipulation

Perception
Laptop 1
Cognition
Robot State
Laptop 2
Action
CAN Bus
Affordable hardware to make teams of robots
practical
4
Ball Manipulation

• A crucial component to soccer performance
• A kicking mechanism is therefore needed
• Based on robot soccer experience the best kickers
  are
• Pneumatic, spring, or motor actuated
• Pneumatics offer simple, efficient solution.
• Tilting platform and soft, uneven surface still
  makes ball manipulation extra challenging

Behaviors must carefully control robot when near
the ball
5
Architecture

• Modular and distributed, built upon our
  hierarchical STP architecture

GUI Offline tools
Debug/Log Server
World Model
Vision
Tracking
Tactics
Adaptive ball/box tracking
Skills
Methods for world state/prediction
Robot Control
Segway Manager
Segway Controller
Motion control, (navigation), raw control
6
Perception

• We have extended CMVision to provide fast color
  object tracking in variable, outdoor lighting
  conditions
• On grassy fields, and some indoor environments,
  robust ball tracking across changing lighting
  conditions

Colorspace Mapping I?(Y,U,V)
Histogram Peak ??peak(Hist(I))
Filters Geometry constraints
Region Growing
Size, shape, expected location Calibrate
expectations
Dot product to 1D Intensity space. Prototype
Vector
Peak detected in histogram. Constraint
parameters
CMVision connected components
7
A Tactic/Skill Example

• ChaseBall for chasing and kicking balls

Sequence complete
Ball not seen in last t seconds
GotoBall Drive to kick location
Kick Playback trained skill
SearchforBall Predict and spin around
Skill State Machine
Observed ball
Ball visible and in kick zone
Motion Control GoTo ltx,y,?gt
Motion Playback
Robot Control Layer
Segway RMP
Segway Manager
8
Skill Recording/Playback

• Developing skills is time consuming and
  non-trivial
• Performance of a skill is highly dependent upon
  robot and environment dynamics
• We have developed a skill recording mechanism to
  ease skill development

XDrive
Tele-operation User interface
Segway Manager
Segway Manager
Record Tactic
Tactic
Motion Recorder
Motion Playback
Recording phase
Playback phase
9
Pickup Segway Soccer

• Cross between Ultimate Frisbee and soccer
• Two teams, two players (Segway RMP or human)
• Simple field markings and human referee

We hope to demonstrate a 2vs2 game at RoboCup
2004, Portugal
10
Summary

• We have developed framework and infrastructure
  for Segway soccer
• Hardware framework and kicking mechanisms
• Control architecture
• Basic perception, tactics, skills and motion
  control
• Future work will focus on
• Extending infrastructure support for development
• Building repertoire of tactics and skills
• Enhancing skill training mechanisms