SayCan

Context

The code of the environment (along with prompt generation and affordance function) was taken from the official notebook release with SayCan: https://github.com/google-research/google-research/blob/master/saycan/SayCan-Robot-Pick-Place.ipynb

We provide in run_pickplace_saycan.py a small script that uses Lamorel (instead of GPT3's API) to score actions. Scoring an action with Lamorel (no matter neither which LLM is used nor whether a distributed setup is used or not) is a single line of code:

lm_server.score([prompt], [possible_actions])

Installation

1. Install required packages: pip install -r requirements.txt

Launch

To launch the example using a single GPU on a local machine:

1. Spawn a process for the RL code:

python -m lamorel_launcher.launch \ 
       --config-path PROJECT_PATH/examples/SayCan/ \
       --config-name PROJECT_PATH/examples/SayCan/local_gpu_config \
       rl_script_args.path=PROJECT_PATH/examples/SayCan/run_pickplace_saycan.py \
       lamorel_args.accelerate_args.machine_rank=0 \
       lamorel_args.llm_args.model_path=PATH_TO_YOUR_LLM

2. Spawn a process for the LLM:

python -m lamorel_launcher.launch \
       --config-path PROJECT_PATH/examples/SayCan/ \
       --config-name PROJECT_PATH/examples/SayCan/local_gpu_config \
       rl_script_args.path=PROJECT_PATH/examples/SayCan/run_pickplace_saycan.py \
       lamorel_args.accelerate_args.machine_rank=1 \
       lamorel_args.llm_args.model_path=PATH_TO_YOUR_LLM