Hybrid Platform

A Gymnasium environment for the Platform domain, featuring a parameterized (hybrid) action space. This environment is widely used for benchmarking Reinforcement Learning algorithms that must handle both discrete action choices and continuous parameters simultaneously.

This is a modern fork of the original gym-platform, fully updated to use the Gymnasium API and modern numpy.

Installation

You can install the package directly from PyPI:

pip install hybrid-platform

Usage

import gymnasium as gym
import gym_platform

# Create the environment
env = gym.make('Platform-v0', render_mode='human')

obs, info = env.reset()

for _ in range(100):
    # Action structure: (discrete_action_index, (param_0, param_1, param_2))
    # where param_X is an array of shape (1,)
    
    # Example: RUN (0) with speed 15.0
    action = (0, ([15.0], [0.0], [0.0]))
    
    obs, reward, terminated, truncated, info = env.step(action)
    
    env.render()

    if terminated or truncated:
        obs, info = env.reset()

env.close()

Action Space

The action space is spaces.Tuple((spaces.Discrete(3), spaces.Tuple((spaces.Box(1), spaces.Box(1), spaces.Box(1))))). It consists of:

1. Discrete(3): The high-level action choice.
   • 0: Run (Move horizontally)
   • 1: Hop (Small jump to clear gaps)
   • 2: Leap (Large jump to clear larger gaps)
2. Tuple(Box(1), Box(1), Box(1)): The continuous parameters for each action.
   • Parameter 0 (Run): Speed, range [0, 30]
   • Parameter 1 (Hop): Power, range [0, 720]
   • Parameter 2 (Leap): Power, range [0, 430]

When taking a step, the environment selects the parameter corresponding to the chosen discrete action index and ignores the others.

Observation Space

The observation space is a spaces.Tuple containing:

1. Box(9,): The state vector (scaled).
   • Basic features: [player_x, player_vx, enemy_x, enemy_dx]
   • Platform features: [platform_width_1, platform_width_2, gap, platform_position, height_diff]
2. Discrete(200): The current time step (used for time awareness).

Original Credits

• Original Platform domain by Warwick Masson et al.
• Original gym-platform implementation by Craig Bester.