Safe Adaptation Agents

Baseline algorithms for the safe-adaptation-gym benchmark.

Install

1. Open a new terminal and git clone the repo.
2. Create a new environment with your favorite environment manager (venv, conda). Make sure to have it set up with python >= 3.8.13.
3. Install dependencies with cd safe-adaptation-agents && pip install ..

Run

Let's reporduce the experiments for the benchmark.

The following command runs the on-policy algorithms:

 python scripts/adaptation_experiment.py --configs defaults multitask on_policy"  --agent <insert agent> --seed <insert seed>

where $\texttt{seed} \in \lbrace1,2, \dots, 10\rbrace$ and $\texttt{agent} \in \lbrace$maml_ppo_lagrangian, rl2_cpo, rarl_cpo$\rbrace$

Similarly, to run the model-based algorithms:

 python scripts/adaptation_experiment.py --configs defaults multitask model_based"  --agent <insert agent> --seed <insert seed>

where $\texttt{agent} \in \lbrace$la_mbda, carl$\rbrace$

To run an agent on a specific task:

 python scripts/no_adaptation_experiment.py --configs defaults no_adaptation" --agent <insert agent> --task <insert task>

where $\texttt{task} \in \lbrace$go_to_goal, dribble_ball, collect, push_box, haul_box, press_buttons, catch_goal, roll_rod$\rbrace$ In our experiments, we used cpo as our agent.

More generally, every parameter in the configs.yaml file can be easily changed when running either scripts/no_adaptation_experiment.py or scripts/adaptation_experiment.py by appending --parameter when running the script.

Benchamrk

We provide the results for our experiments and a jupyter-notebook that summarizes them in the benchmark folder.