NeurIPS 2018: AI for Prosthetics Challenge Solution

A solution based on OpenAI baselines (repo) for 9th place NeurIPS 2018: AI for Prosthetics Challenge

Table of Contents

• Prerequisites
• Training
• Evaluation
• Contributing

Prerequisites

First of all, you should follow the challenge environment's instruction to set up the simulator, details described in: https://github.com/stanfordnmbl/osim-rl/

source activate opensim-rl

Other dependencies is required as follow:

• Ray
• TensorFlow
• mpi4py, cloudpickle, joblib (required by baselines)

Training

To train the model from scratch, you need to comment the --checkpoint-path in run.sh, or type commands as follow:

python -m nips.round2_train \
    --num-timesteps=10000000 \
    --num-steps=128 \
    --num-minibatches=16 \
    --num-cpus=24 --num-casks=8 \
    --save-interval=5 \
    --repeat=2

The baseline code shall be the minimal for this challenge, and we select PPO (paper) as the algorithm.
You should specify your available CPUs and GPUs in the bash script. The training will be deployed distributed by Ray (docs).

Evaluation

To evaluate your model, store your checkpoints file into checkpoints/ and change the arguments in script.

# remember to place your checkpoints in your script file first
./run.sh

As the comparison, our model finally get 9813.224 in the Round 2 of challenge.

Contributing

• wangzhengfei0730
• QPHutu