Search Tree augmented by Language Model (STaLM)

---

This is an official code for STaLM (submitted to IJRR)

Set up

# Setup conda environment
conda env create -f environment.yaml

STaLM makes use of GPT-4 api, so please prepare openai api key. Upon initial learning, you will be prompted to register the openai key

    OpenAI API token is not found at ~/STaLM/LLM/api_token.json. Please enter your OpenAI API token.
    Please enter your OpenAI API token:

Run

To run the STaLM, use

python Simulation/pybullet/experiment_shop.py --override_scenario_num $X --num_episodes $Y --override_num_sims $Z --baseline STaLM

Here, X is the scenario number of the experiment, Y is the number experiments to repeat, and Z is the number of simulation for STaLM to run during WarmStartUCT procedure. To reproduce the result, you can test scenario 1-6 (i.e. X=1-6) with Y=50 and Z=30.

Besides STaLM, you can also test other baselines in the paper by simply giving the name instead of STaLM in the command above. The supported baselines are LLM_MCTS, NO_UCT, MCTS_UCT, Iterative_Replanning, SAYCAN.