Hidden Parameter Recurrent State Space Models (HiP-RSSM)

Pytorch code for ICLR 2022 paper Hidden Parameter Recurrent State Space Models For Changing Dynamics Scenarios. The algorithm learns deep multi task Kalman Filters that can be used in non-stationary environments with changing dynamics.

Dependencies

• torch==1.3.1
• python 3.7
• omegaconf==2.1.1
• hydra-core==1.1.1
• PyYAML==5.3
• wandb==0.10.25
• umap-learn

How to Train

With HiP-RSSM as the working directory execute the python script python experiments/mobileRobot/mobile_robot_hiprssm.py model=default

Datasets

The dataset used here is that of a mobile robot traversing terrain of different slopes as reported in the paper.

For Experimenting With New Datasets

For any dataset with a long timeseries, split them to reasonable local trajectories of length L=2*K, which is fed into the hiprssm model. The first K would used by context encoder to infer latent context and the last K would be used as target set. The concept is very similar to context sets and target sets in Neural Processes or the meta testing procedure used in this reference.

A detailed description for creating training datasets is given in Appendix E. A detailed description for testtime inference procedure is given in Algorithm 1 in the appendix.

How To Run Baselines

With HiP-RSSM as the working directory execute the python script

• LSTM Baseline: python experiments/mobileRobot/mobile_robot_rnn.py model=default_lstm
• GRU Baseline: python experiments/mobileRobot/mobile_robot_rnn.py model=default_gru
• RKN Baseline: python experiments/mobileRobot/mobile_robot_hiprssm.py model=default_rkn