GOHAN

GOHAN means Guided Optimization for High-speed Autonomous Navigation.

This repository is now the GOHAN project root. It contains a telemetry-only reinforcement learning autonomy stack for Autonoma Labs AWSIM Racing Simulator and the Dallara AV-21R. AWSIM runs separately and publishes ROS2 telemetry; GOHAN converts that telemetry into a normalized vector observation, runs a Stable-Baselines3 policy, and publishes steering/throttle/brake commands back through ROS2.

GOHAN does not include AWSIM, clone AWSIM, modify simulator source, use camera perception, or assume CUDA.

Architecture

AWSIM Racing Simulator
  -> ROS2 telemetry
GOHAN Ros2Bridge
  -> telemetry feature extractor
Stable-Baselines3 PPO/SAC-ready policy
  -> action mapper and safety filter
GOHAN Ros2Bridge
  -> ROS2 vehicle command
AWSIM Dallara AV-21R
  -> logging, reward, and training loop

Layout

AG26/
  README.md
  pyproject.toml
  requirements.txt
  configs/
    awsim.yaml
    reward.yaml
    training.yaml
    tracks/default_track.yaml
  scripts/
  src/gohan/
  tests/
  runs/
  ros2_ws/
    src/sim-msgs/
    build/
    install/
    log/

ros2_ws is only for external ROS2 message packages such as autonoma_msgs. The Python package lives directly in this root under src/gohan.

Prerequisites

• Linux Mint 22.x or another ROS2 Jazzy-compatible Linux system.
• ROS2 installed externally, for example /opt/ros/jazzy.
• Python 3.10+.
• AWSIM Racing Simulator installed separately and running in a ROS2-controlled racing scene.
• Autonoma message packages built in ros2_ws.
• CPU-first training. CUDA is not required.

rclpy is intentionally not in requirements.txt; it comes from the sourced ROS2 installation.

Build ROS2 Message Workspace

Run this once from the project root. Use system Python, not Anaconda, so ROS2 CMake finds the right Python packages.

cd /home/osama/Documents/RL/AG26
mkdir -p ros2_ws/src
cd ros2_ws/src

git clone https://github.com/autonomalabs/sim-msgs.git

cd /home/osama/Documents/RL/AG26/ros2_ws
conda deactivate
source /opt/ros/jazzy/setup.bash
colcon build --symlink-install --cmake-args -DPython3_EXECUTABLE=/usr/bin/python3

If a previous configure was polluted by Anaconda, rebuild with:

cd /home/osama/Documents/RL/AG26/ros2_ws
conda deactivate
source /opt/ros/jazzy/setup.bash
colcon build --symlink-install --cmake-clean-cache --cmake-args -DPython3_EXECUTABLE=/usr/bin/python3

Set Up GOHAN Python

Create the Python environment in the root:

cd /home/osama/Documents/RL/AG26
conda deactivate

/usr/bin/python3 -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt
pip install -e .

Every GOHAN Terminal

Run this before any ROS-facing script:

cd /home/osama/Documents/RL/AG26
conda deactivate

source /opt/ros/jazzy/setup.bash
source /home/osama/Documents/RL/AG26/ros2_ws/install/setup.bash
source .venv/bin/activate
export ROS_DOMAIN_ID=0

Start AWSIM

AWSIM does not need to live in this repository. Start it however you installed it, then enter the racing Drive scene with ROS2 control enabled.

export ROS_DOMAIN_ID=0
# launch AWSIM using your installed executable, launcher, or desktop shortcut

Topics will not appear while AWSIM is closed or still outside the running Drive scenario. If GOHAN sees only /parameter_events and /rosout, ROS2 is alive but AWSIM is not visible on that ROS graph.

Manual reset is the default. Automatic reset needs an AWSIM-side reset service or simulator integration; add that later in AWSIMRacingEnv.reset() when a reset API is available.

Inspect ROS2 Topics

ros2 topic list -t
python scripts/inspect_ros_topics.py --config configs/awsim.yaml

Expected high-level racing topics include:

/novatel_top/inspva or /novatel_bottom/inspva
/novatel_top/rawimux or /novatel_bottom/rawimux
/vehicle_data
/powertrain_data
/race_control
/vehicle_inputs
/to_raptor

If a command topic type is unsupported, inspect it:

ros2 topic info /vehicle_inputs

Then update:

src/gohan/ros/message_adapters.py

GOHAN fails clearly instead of publishing an unknown command shape.

Record Telemetry

python scripts/record_telemetry.py --config configs/awsim.yaml --run-name telemetry_test

This verifies that AWSIM telemetry is visible and that the adapters can build GOHAN telemetry rows.

Test Command Publishing

python scripts/drive_manual_test.py --config configs/awsim.yaml --mode slow_forward

Supported modes:

• neutral
• brake
• slow_forward
• sine_steer_low_speed

Run this before training.

Check Environment

python scripts/check_awsim_env.py --config configs/awsim.yaml

This waits for telemetry, creates one observation vector, publishes one neutral command, and exits.

Train PPO

Train only after topic inspection, telemetry recording, environment check, and manual command publishing work.

python scripts/train_ppo.py \
  --config configs/awsim.yaml \
  --training-config configs/training.yaml \
  --reward-config configs/reward.yaml \
  --run-name awsim_ppo_v1 \
  --total-timesteps 100000

Models and logs are written under:

runs/awsim_ppo_v1/

Evaluate

python scripts/evaluate_policy.py \
  --config configs/awsim.yaml \
  --model runs/awsim_ppo_v1/models/final_model.zip \
  --episodes 3

Live Controller

python scripts/run_inference_controller.py \
  --config configs/awsim.yaml \
  --model runs/awsim_ppo_v1/models/final_model.zip

Ctrl+C sends a brake command before shutdown.

Plot Telemetry

python scripts/plot_telemetry.py --input runs/awsim_ppo_v1/telemetry.csv --output runs/awsim_ppo_v1/plots

Observation Vector

GOHAN uses a 20D np.float32 vector clipped to [-1, 1]:

[
  lateral_error_norm,
  heading_error_sin,
  heading_error_cos,
  speed_norm,
  longitudinal_velocity_norm,
  lateral_velocity_norm,
  yaw_rate_norm,
  track_curvature_norm,
  distance_to_left_boundary_norm,
  distance_to_right_boundary_norm,
  lookahead_1_dx_norm,
  lookahead_1_dy_norm,
  lookahead_2_dx_norm,
  lookahead_2_dy_norm,
  lookahead_3_dx_norm,
  lookahead_3_dy_norm,
  previous_steering,
  previous_throttle_brake,
  lap_progress_sin,
  lap_progress_cos
]

The observation is built from ROS2 telemetry and track-relative features, not images.

Action Mapping

The action space is:

gymnasium.spaces.Box(low=-1.0, high=1.0, shape=(2,), dtype=np.float32)

action[0] is steering. action[1] is combined throttle/brake:

if action[1] >= 0:
  throttle = action[1]
  brake = 0
else:
  throttle = 0
  brake = abs(action[1])

Reward

RacingReward combines progress, speed, line tracking, heading alignment, yaw stability, action smoothness, off-track, collision, reverse-progress, and stuck penalties.

Tests

cd /home/osama/Documents/RL/AG26
source .venv/bin/activate
pytest

Unit tests do not require AWSIM or ROS2.

Limitations

• Automatic reset requires AWSIM-side reset support or manual reset.
• Command message types may need adapter updates if AWSIM changes the ROS2 interface.
• The first training goal is stable lap completion, not optimal lap time.
• The default track file is a placeholder; replace it with AWSIM track waypoints or a racing line before serious training.

Roadmap

• v1 direct telemetry-to-control PPO.
• v1.5 reward and curriculum refinement.
• v2 telemetry-guided racing target correction.
• v3 hybrid controller plus RL target optimizer.