Goal Seeking¶

A navigation task where agents move through a grid to reach goal cells. Based on the dynamic decision-making paradigm from Nguyen & Gonzalez (2020).

Description¶

Agents start at spawn points and navigate to goal targets placed on the grid. Each goal has an associated value. The agent receives a reward upon reaching a goal. Supports both single-agent and multi-agent configurations.

Environment Class¶

GoalSeeking extends CoGridEnv and adds:

• Target values — each goal character maps to a reward value.
• Optimal path length — provided per grid configuration for evaluation.
• Multi-agent spawns — separate spawn positions for multi-agent setups.

Objects¶

Goal characters are defined per grid configuration. The grid data maps each character to its reward value.

Rewards¶

The GoalSeekingAgent defines two penalties:

Goal rewards are defined by the target_values mapping in the grid configuration.

Configuration¶

Goal-seeking environments are configured via a grid data file that specifies:

grid_data = {
    "layout": [...],                    # ASCII grid rows
    "values": {"G": 1.0, "X": -0.5},   # character -> reward value
    "optimal_path_length": 8,           # for evaluation metrics
    "ma_spawns": [(2, 1), (3, 1)],      # multi-agent spawn positions
}

The environment is instantiated with a path to this grid data:

from cogrid.envs.goal_seeking.goal_seeking import GoalSeeking

env = GoalSeeking(grid_path="path/to/grid.json", config=config)

Agent API¶

GoalSeekingAgent provides:

• create_inventory_ob() — returns a binary vector of collected target objects.
• inventory_capacity — number of distinct target types.
• step_penalty / collision_penalty — per-step and collision costs.

Spawn Selection¶

The environment supports multiple spawn modes:

• Random spawn — shuffles available spawn points each reset.
• Pre-defined spawn — uses ma_spawns positions for multi-agent setups.
• Generated random spawn — samples from all free spaces (excluding map-specified spawns), controlled by config["env"]["gen_random_spawn"].