논문 제목: “Cafe-MPC: A Cascaded-Fidelity Model Predictive Control Framework with Tuning-Free Whole-Body Control “

저자: He Li and Patrick M. Wensing

<Purpose & Process>

• Dynamics
  • 근거리: Whole body dynamics
  • 원거리: Single Rigid Body
• Multiple-Shooting iLQR →“A unified perspective on multiple shooting in differential dynamic programming” 이전 논문에서 연구한 내용.
  • “A unified perspective on multiple shooting in differential dynamic programming”
    • 초기 추정값에 덜 민감하고 더 안정적인 수렴
    • Optimal Control Problem을 이산화.
    • 각 shooting state을 decision variable로 설정
    • 각 segment에 독립적으로 dynamics integrate
    • forward rollout → trajectory update
    • backward pass → local policy
• VWBC(Value function based whole body control)
  • cafe-mpc의 결과로 얻은 action-value function을 사용해서 tuningX
  • Hard Constraint
• Online, 33~50Hz, 5.3ms average computation time

<Overall framework>

<adaptable features>

• dynamics
  • 나도 가까운곳은 whole body → SRB? ⇒ Rather makes model accuracy low?
• VWBC? → HARD constraint, PD tuning X

<guessing disadvantage(my personal opinion)>

• self collision
• computation efficiency → real time iteration
• footstep planning

<similar paper(may help double check or understand)>