Avinash76812/rl_course_vizdoom_health_gathering_supreme

APPO Agent Playing ViZDoom: doom_health_gathering_supreme 🎮

This is a trained model of an Asynchronous Proximal Policy Optimization (APPO) agent playing the doom_health_gathering_supreme scenario in ViZDoom, trained using Sample Factory 2.0 for Unit 8 (Part 2) of the Hugging Face Deep Reinforcement Learning Course.

Evaluation Results

• Mean Reward: 12.50 +/- 3.20
• Result (mean - std): 9.30 (Passing requirement: >= 5.0)
• Environment: doom_health_gathering_supreme

Scenario Details

In doom_health_gathering_supreme, the agent navigates a rectangular arena with acidic floor terrain that inflicts periodic damage. The agent must collect medkits to heal and prolong survival while navigating obstacles.

• Living Reward: +1 per step
• Death Penalty: -100
• Action Space: Discrete (Turn Left, Turn Right, Move Forward)
• Observation Space: Screen buffer pixels (RGB)

Video Replay

Downloading and Using the Model

After installing Sample Factory and ViZDoom dependencies:

python -m sample_factory.huggingface.load_from_hub -r Avinash76812/rl_course_vizdoom_health_gathering_supreme

To visualize the agent playing:

python -m sample_factory.enjoy --env=doom_health_gathering_supreme --experiment=rl_course_vizdoom_health_gathering_supreme --train_dir=train_dir