Bridging Imagination and Reality for Model-Based Deep Reinforcement Learning (BIRD)

This repository implements the main algorithm of BIRD described in "Bridging Imagination and Reality for Model-Based Deep Reinforcement Learning (NeurIPS 2020)" (arxiv).

Requirements

• run it directly in your python 3 environment with pip:

$ pip install -r requirements.txt

Training

In the folder BIRD/, the following command shows how to train BIRD.

$ python bird.py --logdir ../logdir/dmc_hopper_hop0/ --task dmc_hopper_hop --seed 0 --device 0 --ent_alpha 1e-8 

The default task is Hopper Hop, you can modify the task with --task. Please see DeepMind Control Suite to see all.

Or you can directly run the run.sh to train BIRD, SoftBIRD and Dreamer:

$ bash run.sh

You can also only run BIRD:

$ bash run_bird.sh

Evaluation

We put a checkpoint of BIRD in logdir/dmc_hopper_hop/ to directly test. Use the following commands:

$ python BIRD/bird.py --task dmc_hopper_hop --logdir logdir/dmc_hopper_hop/bird/ 

How to show results

After using run.sh for training, the results for generalization will be directly printed in the screen and the folder logdir/task_name%i/algo_name will be created where task_name is the task name (e.g. dmc_walker_run), %i is the seed, and algo_name is the algorithm name (e.g. BIRD). Or you can set your own logdir to see.

To show the results and gif, run tensorboard with these commands:

$ tensorboard --logdir logdir