Attendees

• Sachin
• Gil
• Caroline
• Ethan
• Eitan
• Brian
• Leila
• Bhaskara
• Peter
• Max, Ben, Mike (PENN)

Agenda

• Coordinate the effort to do motion planning/control for dynamic obstacles and define a set of objectives and roadmap for the same - in particular look at the needs (from a motion planning/control perspective) for people aware navigation
• Define the needs/interface required from people detection/tracking for the planning/control effort

Details

• Planning
  • Planners that can deal with dynamic obstacles
  • Global vs. Local planners
  • Planning times
• Control
  • Local controllers that can deal with dynamic obstacles
  • Modify trajectory rollout to deal with dynamic obstacles
  • Other control schemes
• Representation
  • Define structures for representing dynamic obstacles
    • Position
    • Velocity
    • Uncertainty
    • id
  • Minimally modify current costmap structure to deal with dynamic obstacles
    • Remove dynamic obstacles from incoming sensor stream
    • Deal with dynamic obstacles using forward simulation in trajectory rollout
  • Time varying costmap representations
• Sensing
  • Needs for planning/control
    • position/velocity of dynamic obstacles (can be stubbed out for now)
    • uncertainty - represent expected path (for people)
    • sensor models for obstacles - needed to discriminate static/dynamic obstacles
• Demo scenarios
  • Follow a person down the hallway instead of replanning and turning around
  • Person avoidance - stop and wait/move to the side if person walking towards you

Minutes

Describe motion_planning/Meetings/Minutes 2009-11-18 here.

• Minimal change in current architecture approach
  • Deal with dynamic obstacles separately
  • New trajectory planner plugin takes dynamic world representation
  • World model is a combination of dynamic and static obstacles
  • Is this completely safe - if you think the person is going away from where you will be in the next timestep
    • is there a way to make this more safe?
  • Local planning window may be too small (1.5 seconds)
• Dynamic obstacle tracker returns a list of predicted points in the future instead of returning velocities
  • minimal separation distance - what should this distance be?
    • prevent chattering in this instance
• Global planner can generate a plan to avoid dynamic obstacles (Max)
  • local planner could decide on which dynamic obstacles the global planner needs to be worried about
  • Representation:
    • List of predicted positions in time for each dynamic obstacle
      • x, y + variance to incorporate uncertainty estimates
    • Radius of each dynamic obstacle
    • Allows for multiple trajectories for the same obstacle
      • maybe a person turns into an orthogonal hallway
  • Planning
    • x,y,theta, time - considers the probability of collision as well - if probability is high, those particular positions of robot in time are prohibited
    • finite horizon planning in time
      • longer than x,y,theta lattice planner - planning time within a few seconds
  • For each obstacle - multi-modal representation
    • Radius (Footprint? - need to lay down the footprint) - if the footprint exists, it overrides radius
    • Vector of trajectories
      • Each trajectory - vector of points + 1 probability for each trajectory
        • Each point - x,y,theta,variance (time?)
    • Time - uniform sampling/time for each point in trajectory - could be enforced in API for planner but data structure could still support non-uniform sampling.
• Sensing
  • diff two consecutive sensor views to try and find dynamic obstacles
  • 2D optical flow?
• Action items
  • Create a message to represent the dynamic obstacles (Eitan/Mike(Penn))
  • Perception literature
    • tracking/finding dynamic obstacles
    • optical flow?
    • blob tracker