NavDreams: Social-Navigation World Models

Camera-based robot navigation simulator, world-models, and policies for the paper NavDreams: Towards Camera-Only RL Navigation Among Humans

In NavDreams, we use the world-model/dreamer concept to learn a model which predicts the future, then use this world model to do camera-based robot navigation.

More information available in our paper

Simulator

Our simulator currently contains 7 scenarios. Some of these are based on 2D environments from previous work.

Models and Tools

Model implementations, trained checkpoints, training tools, and plotting tools are made available in this repo. See the How-to section for examples of use.

How-to

• Installation
• Running the simulator
• Training the world-model
• Training the controller
• Testing the controller

Using Docker

First, make sure docker is installed, and, if your computer has a GPU, that nvidia-docker is installed

A pre-built container image is available here. To load it into your docker images:

docker load --input navdreams-docker-image-v2.tar

Then, you can run the container with:

xhost + # allows docker to create UI windows in ubuntu
docker run --rm --gpus all --env="DISPLAY" --net=host --name=n3d_test -it n3d
(inside running container) # python -m navdreams.navrep3danyenv --scenario replica

Not working? Check the troubleshooting guide.

Manual Install

If you already have the typical python/DL dependencies, installing should be as simple as

pip install navdreams

For a more details, refer to this install script

Running the simulated environments

python -m navdreams.navrep3danyenv --scenario replica

See simulator troubleshooting if you encounter issues.

Training the World-Model

Generate the dataset

python make_vae_dataset.py --scope SCR

Train the world-model

python train_gpt.py --dataset SCR

Training the Controller

python train_gym_discrete_navrep3dtrainencodedenv.py --variant SCR --scenario city

and for the end-to-end baseline

python train_gym_discrete_e2enavrep3d_PPO.py

Testing the Controller

Test the controller you trained, (or a downloaded model).

 python test_any.py --model-path \
   "~/navdreams_data/results/models/gym/navrep3dcityencodedenv_2022_02_18__18_26_31_DISCRETE_PPO_GPT_V_ONLY_V64M64_SCR_bestckpt.zip" \
   --n-episodes 100 --build-name "./city.x86_64" --difficulty-mode "hardest" --render True

More

Example Workflow: Evaluating World-model Dreams

---

Troubleshooting Guide

GPU / Docker check (the output should be text information about your GPU):

docker run -it --rm --gpus all pytorch/pytorch:1.9.0-cuda10.2-cudnn7-devel nvidia-smi

Check whether the graphical environment is working. A window with a rotating horse should appear.

sudo docker run --rm --gpus all --env="DISPLAY" --net=host -it n3d bash -c "glmark2"

Running on a laptop

On my T480, the above line

docker run --rm --gpus all --env="DISPLAY" --net=host --name=n3d_test -it n3d

leads to an error due to missing gpu. But when removing the --gpus all flag, things seem to work, except that OpenGL complains about missing drivers, and the simulator sometimes crashes.

libGL error: failed to load driver: i915
libGL error: failed to open /dev/dri/card0: No such file or directory

in this case, replacing --gpus all with --device /dev/dri/ solves these errors, and the simulator can then run smoothly on the laptop with integrated graphics.

xhost +
docker run --rm --device /dev/dri/ --env="DISPLAY" --net=host --name=n3d_test -it n3d

No Screen? No problem.

Are you trying to run this on a computer without a screen (like a cloud server)?
Note that due to using Unity as the simulation backend (MLAgents), a visual environment is required.
It's still possible to do so on cloud servers like AWS (we tested this workflow and it works), but requires a good understanding of the graphics pipeline.
Here's an example on how to start the XServer on a ubuntu server with a Tesla T4 GPU:

# create virtual display with nvidia drivers
# nvidia-xconfig --query-gpu-info
# With headless GPUs (e.g. Tesla T4), which don't have display outputs, remove the --use-display-device=none option
# sudo nvidia-xconfig --busid=PCI:0:30:0 --use-display-device=none --virtual=1280x1024
sudo nvidia-xconfig --busid=PCI:0:30:0 --virtual=1280x1024
sudo Xorg :0 & # unlike startx, this only starts the x server, no DEs
nvidia-smi # check that Xorg is running on the GPU
echo "export DISPLAY=:0" >> ~/.bashrc # applications need to know which display to use

Read here for more details.

If you want to see the virtual display, you can use a vnc viewer, like for example:

sudo apt install -y x11vnc xfce4-session
x11vnc -display :0 -usepw -rfbport 5901
DISPLAY=:0 startxfce4