Fenglong Song

PhD student @EPFL_en | Previously @ETH_en @sjtu1896
Control, optimization, learning, robotics

[Fenglong Song @FenglongS]

@manumerous @breadli428 @zhaomingxie thanks for sharing, Manuel!

[Fenglong Song @FenglongS]

@manumerous @breadli428 @zhaomingxie Also, I saw in many cases the RL policy makes use of knees or other parts of the body pretty well, which is not yet within the reach of MPC (if we refer MPC to solving OCP with SQP or DPP) since contact-implicit MPC is hard. Sampling-based might be more appealing in these cases

[Fenglong Song @FenglongS]

@antoine_leeman @breadli428 Maybe it's possible to learn a residual model to capture the aerodynamic effects using Gaussian Processes and run GP-MPC?😂

[Fenglong Song @FenglongS]

@antoine_leeman @breadli428 Well if the aerodynamics model is differentiable, NLP solver can handle it quite well. But contact results in combinatorial nature, which is mixed-integer programming. Guess MPC can really play a big role in the control pipeline.

[Fenglong Song @FenglongS]

@antoine_leeman I'm always curious how do define robustness in legged locomotion. In locomotion tasks the goal is to maintain the pose of body frame, which seems to be more related to stability. But robust MPC aims to satisfy constraints under uncertainties. I wonder how to relate them😀

[Fenglong Song @FenglongS]

@catachiii Can't believe they made it in just half a year

[Fenglong Song @FenglongS]

@Vlad30847709 all my friends are watching the plane. No one is working at all

[Fenglong Song @FenglongS]

@Vlad30847709 that's true!