A simple C++ implementation of random search for locomotion tasks using MuJoCo.

rs.cpp should work with Ubuntu and macOS.

Operating system specific dependencies:

Install Xcode.

Install ninja:

brew install ninja

sudo apt-get update && sudo apt-get install cmake libgl1-mesa-dev libxinerama-dev libxcursor-dev libxrandr-dev libxi-dev ninja-build clang-12

git clone https://github.com/thowell/rs.cpp

1. Change directory:

cd rs.cpp

2. Create and change to build directory:

mkdir build
cd build

3. Configure:

cmake .. -DCMAKE_BUILD_TYPE:STRING=Release -G Ninja

cmake .. -DCMAKE_BUILD_TYPE:STRING=Release -G Ninja -DCMAKE_C_COMPILER:STRING=clang-12 -DCMAKE_CXX_COMPILER:STRING=clang++-12

4. Build

cmake --build . --config=Release

VSCode and 2 of its extensions (CMake Tools and C/C++) can simplify the build process.

1. Open the cloned directory rs.cpp.
2. Configure the project with CMake (a pop-up should appear in VSCode)
3. Set compiler to clang-12.
4. Build and run the rs target in "release" mode (VSCode defaults to "debug").

Train cheetah in ~10 seconds using 20 threads with Intel Core i9-14900K CPU and Ubuntu 22.04.4 LTS (~20 seconds using 10 threads on Apple M1 Pro).

From the build/ directory, run:

./rs --env cheetah --search --visualize --checkpoint cheetah

The saved policy can be visualized:

./rs --env cheetah --load cheetah --visualize

Environments available:

• Ant
  • based on ant_v5
  • modified solver settings
  • only contact between feet and floor
  • no rewards or observations dependent on contact forces
• Cheetah
  • based on half_cheetah_v5
  • modified solver settings
• Humanoid
  • based on humanoid_v5
  • modified solver settings
  • only contact between feet and floor
  • no rewards or observations dependent on contact forces
• Walker
  • based on walker2d_v5
  • modified solver settings
  • only contact between feet and floor

Note: run multiple times to find good policies.

Run from build/:

Search:

./rs --env ant --search

Visualize policy checkpoint:

./rs --env ant --load pretrained/ant_2063_239 --visualize

Search:

./rs --env cheetah --search

Visualize policy checkpoint:

./rs --env cheetah --load pretrained/cheetah_3750_239 --visualize

Search:

./rs --env humanoid --search

Visualize policy checkpoint:

./rs --env humanoid --load pretrained/humanoid_2291_1931 --visualize

Search:

./rs --env walker --search

Visualize policy checkpoint:

./rs --env walker --load pretrained/walker_5619_41 --visualize

Setup:

• --env: ant, cheetah, humanoid, walker
• --search: run random search to improve policy
• --checkpoint: filename in checkpoint/ to save policy
• --load: provide string in checkpoint/ directory to load policy from checkpoint
• --visualize: visualize policy
• --num_threads: number of threads / parallel workers

Search settings:

• --nsample: number of random directions to sample
• --ntop: number of random directions to use for policy update
• --niter: number of policy updates
• --neval: number of policy evaluations during search
• --nhorizon_search: number of environment steps during policy improvement
• --nhorizon_eval: number of environment steps during policy evaluation
• --random_step: step size for random direction during policy perturbation
• --update_step: step size for policy update during policy improvement
• --nenveval: number of environments for policy evaluation
• --reward_shift: subtract baseline from per-timestep reward

• This implementation is not deterministic. You may need to run the search multiple times to find a good policy.
• The environments are based on the v5 MuJoCo Gym environments but may not be exact in all details.
• The search settings are based on Simple random search provides a competitive approach to reinforcement learning: Table 9 but may not be exact in all details either.