Iterative LQG for a couple of MuJoCo models. Find a write-up HERE.
A few iterations before convergence is shown here. Hopper in particular does some crazy stuff, but it will get there (no wonder why evolution wiped out these guys). This whole process below is considered as a signle MPC run. You can set contact solref[0] to a slightly higher number in the model, for more smooth dynamics and hence better derivative behaviors.
- Iterative LQR algorithm, based on this paper:
- Improved Regularization
- Backtrack Linesearch
- Finite difference calculation of derivatives, in parallel
- The Levenberg-Marquardt heuristic
- Hessian approximation from gradients
- MPC mode
Modify the Makefile by specifying MJKEY and MUJOCO_PATH, make the executables with make,
and run with ./bin/base ./res/inverted_pendulum.xml.
- Make it faster of course
- Extend to walking robots
- Use proposed cost function from the paper
Big thanks to Taylor Apgar, from Dynamic Robotics Laboratory at Oregon State University.