ALFWorld Episode Reward¶

`alfworld_episode_reward(context: Context) -> Dict[str, Any]` `async` ¶

Simple ALFWorld episode reward that checks if the episode is done. Uses the same rollout resource as the alfworld tools (context.acquire_resource).

Source code in src/agentfly/rewards/alfworld_reward.py

@reward(name="alfworld_episode_reward")
async def alfworld_episode_reward(context: Context) -> Dict[str, Any]:
    """
    Simple ALFWorld episode reward that checks if the episode is done.
    Uses the same rollout resource as the alfworld tools (context.acquire_resource).
    """
    env = await context.acquire_resource(spec=ALFWorldSpec, scope="global", backend="local")
    _obs, reward_val, _done, _info = await env.step("")
    if reward_val is None:
        reward_val = 0.0

    return {
        "reward": reward_val,
    }

Description¶

Evaluates agent performance in ALFWorld tasks by checking the episode completion status and reward from the environment state. Returns a dict with reward (the environment's current reward value).

Decorator Configuration: - name: "alfworld_episode_reward" - resource_spec: ALFWorldSpec - backend: "local"

Technical Details¶

Implementation: - Steps environment with empty action to get current state - Extracts reward value from environment response - Handles None reward values by defaulting to 0.0 - Provides debug output for reward values

Use Cases: - Evaluating task completion in household environments - Training agents on multi-step instruction following - Measuring progress in text-based interactive environments

Example Usage:

from agentfly.core import Context
from agentfly.rewards.alfworld_reward import alfworld_episode_reward

# Inside a rollout, `context` is injected and passed through to rewards.
result = await alfworld_episode_reward(
    prediction="take apple",
    context=context,
)
print(result)  # {"reward": 0.0} or {"reward": 1.0} if task completed

Environment Integration: - Requires active ALFWorld environment instance - Uses environment's internal reward mechanism - Suitable for episodic task evaluation

ALFWorld Episode Reward¶