[기초] TAMP: Task and Motion Planning 소개

온전히 제 이해를 위해 작성한 글입니다. 

• TAMP: Task and motion planning - a sequence of low-level primitive planning
  •  
  • Task Planning: 어떤 action을 취할지에 대한 plannning
  • Motion Planning: 해당 action이 정해졌다면 어떤 motion을 이용할 것인지에 대한 Planning

해결 방안

• Pure Planning: 
  • Model based Simulation, Learning a planner: Carefully designed models + Planner
  • Complete / Inefficient
  • 종류
    • Offline planning: optimal
    • Online planning: MCTS
    • Meta-planning: Learning to guide planning
      1. Use a relaxed problem representation as an input to a planner neural network
      2. Use planning Learning to solve the relaxed problem
      3. Use the output as a planning guidance 
      4. → Geometric predicate로 정의 → State로 묶어서 → GNN

• Pure Learning
  • Learning a policy ex. Deep RL algorithms 
  • Fast/Fragile(hard to apply on long term)/Data-hungry
  • 종류
    
    • Model-based
    • Model-free

Pure online planning 모델 종류

• MDP
• POMDP
  • POMCP
  • DESPOT
• TAMP
• MCTS
  • MCTS+RL: Alphago