google-deepmind
Question about different performance
Hi, I am using this repository now and also learning the source code, but I have a question.
I used the "python mujoco_mpc/demos/agent/cartpole_gui.py", and the Shadow task looks really good. But when I ran the "mujoco_mpc-main\python\mujoco_mpc\demos\predictive_sampling\particle.py" code, it almost doesn't work. Is this because in this demo, we don't sample enough to find the good actions? Normally, how many iterations it needs and can the Horizon be set longer like 1s?
Really appreciate your response in advance!
Will you please provide additional detail for what 'almost doesn't work' means in this context? Thanks!
An important different between the two demos is that cartpole_gui.py is calling the C++ code via the Python API while particle.py is calling a Python implementation of predictive sampling. Given the difference in programming language implementations it is expected that there is a performance difference in terms of wall clock time.
