DRL quadruped hybrid navigation

Installation

It is recommended to install IsaacSim's 5.1.0 pre-built binaries, following Nvidia's official guide. If no binary installation is detected, the setup script will use isaacsim pip package, which lacks VsCode IntelliSense support

$ ./setup.sh
$ source .venv/bin/activate

Dataset setup

Clone huggingface repo (25GB of disk required)

# setup git-xet
$ curl -sSfL https://hf.co/git-xet/install.sh | sh

$ git clone https://huggingface.co/datasets/spatialverse/InteriorAgent

# preprocess the dataset
$ python3 src/preprocessing/preprocess_interioragent.py 

Set the dataset_folder parameter in dataset_cfg.yaml to the InteriorAgent repo folder.

Weights & Biases setup (optional)

Add the wandb username and project name into a .env file:

$ echo -e "WANDB_USERNAME=<user> \nWANDB_PROJECT=<project>" > .env

Usage

Train a simple policy that navigates in empty environments

$ python3 train_policy.py --num_envs=128 --max_iterations=500 --task="go2_lidar_empty" --headless

The available tasks are:

- go2_lidar_empty
- go2_lidar_full
- go2_vision_empty
- go2_vision_full
- go2_depth_full

where the *_full tasks build the entire scene, whereas *_empty ones are just an empty plane.

The complete MDP configurations are defined in the src/tasks/go2_*.py files.