Skip to content

README.md

Latest commit

 

History

History
99 lines (69 loc) · 4.05 KB

README.md

File metadata and controls

99 lines (69 loc) · 4.05 KB

Arena-Rosnav

A modular ROS 2 (Jazzy) platform for researching and benchmarking autonomous robot navigation in 2D and 3D simulated environments. It supports classical planners (Nav2), deep-RL planners (rosnav_rl), and a variety of simulators (Gazebo, Isaac Sim).

Installation

Prerequisites: Docker installation with nvidia-container-toolkit for GPU support. Current user must be in group docker. Afterwards, run the following commands to install Arena:

Basic Installation

curl https://raw.githubusercontent.com/voshch/Arena/jazzy/install.sh > install.sh
bash install.sh

and follow the prompts. This will create a ROS 2 workspace at your target location and instruct you how to proceed (yellow text).

Optional Features

cd ~/arena_ws # replace with your actual workspace path
source arena
arena feature isaac install # optional
arena feature gazebo install # optional
arena feature training install # optional
arena feature vllm install # optional: local LLM backend

We recommend installing at least one simulator.

vllm

Runs a local vLLM server plus a LiteLLM proxy that speaks the Gemini API, so GPT consumers in task_generator transparently hit local inference instead of Google. Defaults target an 11 GB 2080 Ti (Qwen3-0.6B, 40% GPU util).

Tune via _meta/docker/features/vllm/config.yaml:

key default purpose
model Qwen/Qwen3-0.6B HF model id
gpu_memory_utilization 0.4 fraction of VRAM vllm may claim
max_model_len 4096 context window
port / proxy_port 8000 / 4000 vllm / LiteLLM ports

After editing, re-run arena feature vllm update to recreate the container. The container will start automatically on source and continue running in the background. To free up GPU memory, stop it with arena feature docker stop.

Usage

cd ~/arena_ws # replace with your actual workspace path
source arena
arena launch sim:=isaac                          # Isaac Sim
arena launch local_planner:=rosnav_rl agent_name:=<your_agent>  # DRL planner
arena launch sim:=gazebo local_planner:=rosnav_rl env_n:=2 train_config:=<path to config.yaml> # DRL training

DRL quick-start

Place your trained agent folder inside Arena/arena_training/agents/<agent_name>/ (must contain training_config.yaml and best_model.zip), then launch with local_planner:=rosnav_rl agent_name:=<agent_name>. Refer to the arena_training for training instructions.

Development

Linting

Linting is handled by Ruff, driven by pre-commit. Config lives in root pyproject.toml; the hook pin is in .pre-commit-config.yaml. Auto-formatting is intentionally not enforced.

One-time setup:

pip install pre-commit
pre-commit install

Everyday use: hooks run automatically on git commit against staged files. To run manually:

pre-commit run            # staged files only
pre-commit run -a         # entire repo
ruff check .              # check without pre-commit

If the hook auto-fixes something, the commit is aborted and the fixes are left unstaged — git add and re-commit.

CI

.github/workflows/lint.yml runs the same pre-commit hooks on every push to jazzy and every pull request targeting it. The GH check uses the exact config and hook pins from .pre-commit-config.yaml, so local and CI never drift. Make the check required in branch protection to block merges on lint failures.

Bump the Ruff version with pre-commit autoupdate.

Troubleshooting

Unknown runtime specified 'nvidia'

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker
sudo nvidia-ctk runtime configure --runtime=containerd
sudo systemctl restart containerd