I think this was already covered when I read Erez’ thesis, but it was on my kindle, so here goes

- Combines offline trajectory optimization and online model predictive control (MPC)
- The method for the hopper domain
- Use offline optimization to find the limit cycle for the task, and then compute a locally quadratic estimation of the value around the limit cycle. This quadratic estimation is used as the terminal cost for online MPC
- The approach Infinite-horizon Model Predictive Control
- Was previously only applied to small domains, first time applied to periodic domains

**They claim online optimization is ineffective in domains with contacts because of myopic behavior stemming from the limited horizon**
**I remember reading this in the thesis, it is an artifact of the approach they use and not a fundamental limitation**
- May have to do with the fact that they plan from the end of the sequence to the front?

- Infinite horizon planning can be used in domains with terminal states or limit cycles
- Planning falls into 2 categories:
- Simultaneous methods explicitly represent the trajectory in state space, treating the dynamics as constraints
- Sequential methods represent only the control sequence and use integration to evaluate the trajectory

- Apparently shooting algorithms attempt to solve BVP problems?
- They use the simultaneous (state-based policy) to plan the limit cycle offline, and sequential planning for the online part
**They build a model that uses inverse dynamics**
- They need that so that they can just define the trajectory through the state-space desired and then derive the actions from that

### Like this:

Like Loading...

*Related*