Environment API

Core Methods

from pogema import pogema_v0, GridConfig

env = pogema_v0(GridConfig(num_agents=2, size=8, seed=42))

# Reset environment
obs, info = env.reset(seed=None)

# Take a step (actions: list of ints, length = num_agents)
obs, reward, terminated, truncated, info = env.step(env.sample_actions())

# Sample random valid actions
actions = env.sample_actions()  # np.ndarray of shape (num_agents,)

# Get number of agents
n = env.get_num_agents()

# Render to terminal
env.render()

State Access Methods

# Agent and target positions (list of [x, y])
agents = env.get_agents_xy()
targets = env.get_targets_xy()

# Positions relative to initial placement
agents_rel = env.get_agents_xy_relative()
targets_rel = env.get_targets_xy_relative()

# Obstacle grid (numpy array, 1 = obstacle)
obstacles = env.get_obstacles()

# Full state as array or dict
state = env.get_state()
state_dict = env.get_state(as_dict=True)

# Per-agent goal status from last step
was_on_goal = env.was_on_goal  # list of bool

Spaces

env.observation_space  # gymnasium.spaces.Box or Dict
env.action_space       # gymnasium.spaces.Discrete(5)

Method Parameters

reset()

Parameter	Type	Description
seed	int | None	Override environment seed for this reset
options	dict | None	Additional options (unused currently)

step()

Parameter	Type	Description
action	list[int]	One action per agent (0=wait, 1=up, 2=down, 3=left, 4=right)

get_agents_xy() / get_targets_xy()

Parameter	Type	Default	Description
only_active	bool	False	Only return active (non-finished) agents
ignore_borders	bool	False	Return positions without border offset