Setting up f1tenth_gym_ros with 4 cars

IROS 2026 races up to four cars on the track at the same time, so it is worth practising against more than one opponent well before you travel. The simulator has moved to ROS 2 Jazzy: multi-car support is on the dev-jazzy branch of f1tenth_gym_ros, which runs the dev-jax branch of f1tenth_gym, the JAX rewrite of the gym. Still on Ubuntu 22.04 / ROS 2 Humble? Use the dev-humble branches of both repositories instead; everything below applies, except that the launch argument there is still spelled num_agent.

Installing

The bridge needs Ubuntu 24.04 with ROS 2 Jazzy (natively, recommended) or Docker with Foxglove in your browser. The gym lives inside the bridge’s clone and is installed into a virtual environment that can see the system ROS packages:

mkdir -p ~/sim_ws/src && cd ~/sim_ws/src
git clone -b dev-jazzy https://github.com/f1tenth/f1tenth_gym_ros.git
git clone -b dev-jax https://github.com/f1tenth/f1tenth_gym.git f1tenth_gym_ros/f1tenth_gym
python3 -m venv --system-site-packages ~/sim_ws/.venv && source ~/sim_ws/.venv/bin/activate
pip install -e ~/sim_ws/src/f1tenth_gym_ros/f1tenth_gym
cd ~/sim_ws && rosdep install -i --from-path src --rosdistro jazzy -y && colcon build --symlink-install

The README has the complete step-by-step list (system packages, the Docker route, a FAQ for the usual errors).

Number of cars

Set the number of cars with num_agents, either in config/sim.yaml or directly at launch:

ros2 launch f1tenth_gym_ros gym_bridge_launch.py num_agents:=4

Note the plural: the old num_agent spelling is refused at launch with a message rather than silently running a single car.

Driving the opponents

The first car is the ego, the rest are opponents. Each opponent gets its own topics (/opp_drive, /opp_drive2, /opp_drive3, and likewise /opp_scan, /opp_scan2, …, /opp_racecar/odom, /opp_racecar2/odom, …), and a car only moves when you publish to its drive topic — an opponent you never command will simply sit on the grid. Every car also publishes collision, lap_count and lap_time topics under its namespace, so you can score a practice race without any extra tooling.

Start poses

Every car needs a start pose in the config (sx1/sy1/stheta1, …); the bridge refuses to start and names the missing parameters otherwise. The shipped sim.yaml lines up eight cars on the Levine map, so anything up to num_agents:=8 works out of the box. Poses refer to base_link, the rear axle, as on the real car.

Visualisation

Foxglove opens automatically in the browser (open_foxglove:=false to stop that, e.g. in a container). The shipped layout config/foxglove/gym_bridge_foxglove.json draws all eight cars plus a lap counter and lap timer for the ego. RViz still works with config/rviz/gym_bridge.rviz.

Training with reinforcement learning

For learning-based approaches, use f1tenth_gym on dev-jax directly, without ROS. It is natively multi-agent (step takes a (num_agents, 2) action array), runs its physics in JAX and batches many environments on a GPU, and ships a PureJaxRL-style PPO example (examples/jax_ppo_training.py). Gymnasium wrappers flatten it to the single-agent Box spaces that Stable-Baselines3 and CleanRL expect, with an optional SBX adapter.

git clone -b dev-jax https://github.com/f1tenth/f1tenth_gym.git && cd f1tenth_gym
uv sync --frozen            # add --no-group gpu on a machine without CUDA
uv run python examples/waypoint_follow.py

The RL guide covers the batch API and the wrappers. Note that the ROS bridge pins its physics to the CPU; a GPU only matters for your own training code.

Full documentation

Bugs or problems with any of them: please open a GitHub issue on the repository concerned.