jtatman's picture
Upload folder using huggingface_hub
723d096 verified
|
raw
history blame
2.91 kB

Robot Volleyball

A simple minigame arcade-style Volleyball example with AI vs AI and Human VS AI modes.

Observations:

  • Position of the ball in robot's local reference,
  • Position of the ball in the robot's goal's local reference (to tell on what side of the field the ball is in),
  • Velocity of the robot in robot's local reference,
  • Velocity of the ball in robot's local reference,
  • Whether the jump sensor is colliding (which means that the robot can jump),
  • Whether the robot is serving,
  • Whether the ball has been served,
  • Hit count of the ball in a row (hitting the ball more than 2 times in a row by the same robot causes a fault),
  • How many steps has passed without hitting the ball (only counted if serving, there is a time limit in that case).

Action space:

func get_action_space() -> Dictionary:
    return {
        "jump": {"size": 1, "action_type": "continuous"},
        "movement": {"size": 1, "action_type": "continuous"}
    }

Rewards:

  • Positive reward for hitting the ball once when serving,
  • Negative reward if the same robot hits the ball more than 2 times in a row,
  • Negative reward if the ball hits the robots own goal

Game over / episode end conditions:

In infinite game mode or training mode, there are no specified end conditions. It's possible to disable these modes in the GameScene node in which case a winner will be announced, and the scores will be restarted, after a certain amount of points is reached.

Running inference:

AI vs AI

Open the scene res://scenes/testing_scenes/ai_vs_ai.tscn in Godot Editor, and press F6 or click on Run Current Scene.

Human vs AI

To play VS the AI, open the scene res://scenes/testing_scenes/human_vs_ai.tscn in Godot Editor, and press F6 or click on Run Current Scene.

Controls (you can adjust them in Project Settings in Godot Editor):

Volleyball Controls

Training:

The default scene res://scenes/training_scene/training_scene.tscn can be used for training.

These were the parameters used to train the included onnx file (they can be applied by modifying stable_baselines3_example.py):

    policy_kwargs = dict(log_std_init=log(1.0))
    model: PPO = PPO("MultiInputPolicy", env, verbose=1, n_epochs=10, learning_rate=0.0003, clip_range=0.2, ent_coef=0.0085, n_steps=128, batch_size=160, policy_kwargs=policy_kwargs, tensorboard_log=args.experiment_dir)

The arguments provided to the example for training were (feel free to adjust these):

--timesteps=6_500_000
--n_parallel=5
--speedup=15
--env_path=[write the path to exported exe file here or remove this and n_parallel above for in-editor training]
--onnx_export_path=volleyball.onnx