Code for paper Online Learning and Control of Dynamical Systems from Sensory Input (NeurIPS 2021)

Files

The figures folder contains subfolders named after figures from the main submission and from the supplementary material. Each subfolder from fig1 to fig5 contains one GT video sequence, and predicted video sequences with the different models presented in the paper.

Code

0. Managing environments

For training, evaluation and control, use odc.

cd online_dynamics_control
conda env create -f odc_environment.yml --name odc
conda activate odc
pip install -e .

An unactuated pendulum dataset can be downloaded from: https://mybox.inria.fr/f/27c2413966a844ea8506/?dl=1 .
An actuated pendulum dataset can be downloaded from: https://mybox.inria.fr/f/564ebf586a8c45148202/?dl=1 .
To generate more systems, see section 2.

1. Models

1.1 Train

First, make sure your dataset folder (the folder containing train and test folders) is in the datasets folder.

Then, run:

conda activate odc
cd odc
python train.py --data_path datasets/pendulum_v_low_eval_64_64_bw --lr 1e-3 --ae_out_dim 8 --dataset_fraction 1 \
--delayed_dmd 1 --history 2 --a_method prox --online_update 0 --exp_name test_run_
cd ..

Checkpoints will be saved in odc/runs/

1.2 Keep only checkpoints with lowest loss

bash clean_runs.sh

1.3 Evaluate

cd odc
python evaluate_model.py --runs_dir models/ --exp_name simple_low_eval_prox_2021_5_23_21_54 \ 
--eval 1 --sample_duration 300 --metric_prefix sd=300_ --data_path datasets/pendulum_v_low_eval_64_64_bw
cd ..

Complete test_run_ with complete name with date

1.4 Control

This will generate Figure 7 of the supplementary material

cd odc
python control.py
cd ..

To generate Figure 6 from the main submission, run:

python odc/control.py --exp_path models/simple_l=0.6_control_sd=200_nptsA=150_2021_5_19_0_41 \ 
--model_path models/simple_l=0.6_control_sd=200_nptsA=150_2021_5_19_0_41/model_loss_8.262357005150989e-05.pth \ 
--data_path datasets/pendulum_v_l=0.6_control_sd=200_a=0.167_64_64_bw \ 
--system pendulum --video_idx 1301 --idx_init 151 --idx_final 178

2. Generate data

To generate a cartpole dataset, use the pino environment.

cd online_dynamics_control
conda env create -f pino_environment.yml --name pino
conda activate pino
pip install -e .

And run:

conda activate pino
bash generate_data.sh

or edit it for other systems.