Implementation of some core elements of "Learning robust perceptive locomotion for quadrupedal robots in the wild" [Paper]
This repository includes implementation of two elements.
- Student policy network
- Heightmap noise generator
Student policy network is composed of belief encoder and belief decoder to appropriately fuse both proprioceptive and exteroceptive sensor data. It is implemented in Python. Privilege information decoder, included in the paper, is excluded because they were not that critical in our experiement.
Heightmap noise generator is composed of three noise models to handle errors available in real-world use cases due to depth camera noise, state estimation error/drift etc. It is implemented in C++ because the Raisim simulator that we are actively using implements environments in C++ for fast simulation.
- numpy
- pytorch
- ruamel.yaml
- Student policy network
cd model
python main.py
- Heightmap noise generator
cd noise_generator
mkdir build && cd build
cmake ..
make
# After build is finished
./noise_example