Locked Gem is a simple grid-based Reinforcement Learning environment with a sparse reward and a required sequence of actions.
The agent must first activate a pressure plate to unlock a door before it can reach the gem. The agent only observes the environment as an image (pixels), with no symbolic state information.
- Go to the pressure plate
- Unlock the door
- Go through the door
- Collect the gem
Collecting the gem ends the episode and gives a reward of +1.
| Event | Reward |
|---|---|
| Collect gem | +1.0 |
| Anything else | 0.0 |
The reward is intentionally sparse.
- The observation is a single RGB image of the grid
- No access to positions, coordinates, or symbolic information
- The agent must learn everything from pixels alone
Discrete movement actions:
- Up
- Down
- Left
- Right
The episode ends when:
- The gem is collected, or
- A maximum number of steps is reached
Locked Gem is designed to be challenging despite its simple rules:
- Sparse terminal reward
- Long sequence of required actions
- Subgoal dependency (plate → door → gem)
- Pixel-only observations
A key goal of this environment is generalization.
Agents are expected to:
- Train on a set of environment configurations (layouts)
- Generalize to new, unseen configurations
This requires learning what to do (activate plate, pass door, collect gem), not memorizing paths or positions.
import gymnasium
import locked_gem_env
env = gymnasium.make(
"locked_gem_env/LockedGem-v0",
size=12,
render_mode="human",
max_episode_steps = 50
)
obs, info = env.reset(seed=42)
done = False
total_reward = 0.0
while not done:
action = env.action_space.sample()
obs, reward, terminated, truncated, info = env.step(action)
done = terminated or truncated
total_reward += reward
print("Episode reward:", total_reward)
env.close()The environment supports RGB rendering, as shown in the example above.
MIT License (or specify otherwise).
Pedro Mota