The related paper had been accepted bt ICRA2025! Now we refactor here's code toward v.1.0 release Fossen' physics model based multiple platform simulator including dynamics, reinforcement learning script, robot models, actuator models.
eeUVsim has task oriented RL training script with randomize and repeating function
The training environment is impelemted based on ros2 pub-sub method as actual robot software basically has. Then it is easy to transfer to real robot from simulator.
| Ubuntu version | ROS2 Version | Comment |
|---|---|---|
| 22.04 | Humble | Recommnded |
| 20.04 | Galactic |
| Package | Version | Comment |
|---|---|---|
| Python | 3.10 | Basically should work with ubuntu default version |
| ROS2 | (Above) | |
| Gazebo | 11 (Classic) | It is only used for visualization without physics |
| Gym | 0.26.2 | It used for the reinforcement learing environment |
| Stable-baselines3 | 2.3.2 | Used to train reinforcement learning agent |
| sb3_contrib | 2.3.0 | Used for LSTM-PPO training |
In UCAT case, you can launch fundamental scripts (motion controller, dynamics, robot model) with this command.
ros2 launch eeuv_sim spawn_UCAT.launch.py
When you want to train surfacing controller, you can run
ros2 run eeuv_sim RL_UCAT.py -o
Then iterative training will start.