One Shot Damage Diagnosis

This is the code for a portion of my B.Tech Thesis Project Deep Reinforcement Learning for Stability and Safe Adaptation in Damaged Robots. It can diagnoze damage in any locomatory OpenAI gym agent using only one rollout of 20 timesteps.

Requirements

1. Keras
2. Tensorflow
3. OpenAI gym and Mujoco (See installation instructions here)
4. Joblib

How to use

First collect samples of damaged robot data using

python sampler.py experts/Ant-v1.pkl Ant-v1 --num_rollouts 2000

(Note that this step is paraellized over multiple threads. I have written this code for 4 threads, it can easily be scaled up for clusters having large number of available threads.)

Then load the pickled data and train the LSTM network

python rnn_train.py data_pickles/Ant-v1_4joints20diff102type1.dict -s saved_models/myclean.h5 -e 50

Generate some test data again using sampler and run testing network

python rnn_test.py -m saved_models/my_modelant4jointsday32_eff_div_2type.h5 -d data_pickles/Ant-v1_4joints20diff1002type1.dict

This work is still in progress. Feel free to contact me if you are interested in this kind of architecture or want to discuss any ideas.