Efficient Planning in a Compact Latent Action Space

Unofficial Knet.jl implementation of the "Efficient Planning in a Compact Latent Action Space".

This version is implemented by Enes Erciyes for the Koç University Comp 541 Course. You can find the original implementation in https://github.com/ZhengyaoJiang/latentplan.

Tech report of the reproduction effort can be found here.

Setting up D4RL and PyCall

• Download and place MuJoCo 2.10 in ~/.mujoco/mujoco210.

• Add this path to LD_LIBRARY_PATH in your shell init script.

• Add PyCall and Conda in your Julia environment.

• Inside a Julia REPL, set ENV["PYTHON"] = "" and run using PyCall. This will set up a conda environment called conda_jl.

• Inside a Julia REPL, run Conda.add("glew"; channel="conda-forge") and Conda.add("mesalib"; channel="conda-forge").

• Then, run Conda.pip_interop(true) to be able to install pip dependencies.

• Install the pip dependencies using:

Conda.pip("install", ["mujoco-py==2.1.2.14", "git+https://github.com/JannerM/d4rl.git@c3dd04da02acbf4de6cbaa1141deb4f958f03ca9", "dm_control", "git+https://github.com/aravindr93/mjrl@3871d93763d3b49c4741e6daeaebbc605fe140dc"])

• Outside the Julia REPL, run

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/conda/pkgs/[mesalib-pkg-dir]/lib
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/conda/pkgs/[zstd-pkg-dir]/lib

• Check if the setup is successful by importing d4rl:

using PyCall
d4rl = pyimport("d4rl")

NOTE: Flow and CARLA environments are not used here, therefore they are not included in the setup.

Using pretrained checkpoints

Download the main model, prior model and dataset config from these links:

Main model: Drive link

Prior model: Drive link

Dataset config: Drive link

Make the following directory and put the files there:

~/logs_julia/hopper-medium-replay-v2/T-1-1/

Run the following command:

for i in {2..20};
do
    julia --project=.. plan.jl --dataset hopper-medium-replay-v2 --exp_name T-1-1 --suffix $i --n_expand 4 --beam_width 64 --horizon 15
done