genesis_manipulation

About

A motion planning example for a Panda arm using Genesis and reinforcement learning.

setup

git clone https://github.com/TakashiSato/genesis_manipulation.git ~/genesis_ws/genesis_manipulation
cd ~/genesis_ws/genesis_manipulation/
bash scripts/setup.bash

train

cd ~/genesis_ws/genesis_manipulation/
source ~/genesis_ws/.venv/bin/activate
python3 src/panda_train.py --max_iterations 100 -B 4096

• -B is the number of parallel environments

train with viewer

• append --show_viewer to the train command
• NOTE
  • visualize only the first environment
  • train will be slower but you can see the training process

python3 panda_train.py --max_iterations 100 -B 4096 --show_viewer

train with viewer and visualize parallel

• append --show_viewer and --show_parallel to the train command
• NOTE
  • train will be more slower than the above command

python3 panda_train.py --max_iterations 100 -B 128 --show_viewer --show_parallel

visualize training states with tensorboard

1. execute panda_train.py
2. execute tensorboard in another terminal

source ~/genesis_ws/.venv/bin/activate
tensorboard --logdir ~/genesis_ws/genesis_manipulation/logs

3. open http://localhost:6006/ in your browser

evaluate trained model

cd ~/genesis_ws/genesis_manipulation/
source ~/genesis_ws/.venv/bin/activate
python3 src/panda_eval.py --ckpt 100