Hide-And-Seek Engine (SAR Extension)

High-performance OpenMP + pybind11 grid-world simulator for heterogeneous Search and Rescue (SAR), with:

• CTDE-ready tensors (C x H x W) for CNN extractors
• Hybrid action space (move + radio)
• PettingZoo parallel API adapter
• Local/POV rendering utilities

Install

pip install -e .

Optional rendering/input dependencies:

pip install pygame pillow pettingzoo imageio

Level File Formats

test_level/tiles.json

Either list or name->object map. Each tile supports:

• rgb: [r, g, b]
• altitude: float
• supports_walking: bool
• supports_flying: bool
• supports_aquatic: bool
• blocking: bool

Movement semantics:

• Agent can enter tile when it matches at least one supported transport mode.
• If transport does not match:
  • blocking=true: tile behaves like wall (entry denied, agent not stuck)
  • blocking=false: agent can enter but becomes stuck

test_level/agents.json

Either list or name->object map. Each agent supports:

• flying, aqueous, walking
• altitude_min, altitude_max
• base_speed, base_view, battery, deployment_delay
• rgb
• terrain_speed dictionary by tile name
• start ([y, x], supports normalized [0..1] or map coords)

test_level/survivors.json

Either list or name->object map. Each survivor supports:

• allowed_savers: list of agent names
• moves: bool
• rgb (optional)
• start (optional)

test_level/level.png

PNG map where every pixel is matched to nearest tile rgb in tiles.json.

Core Environment Usage

from hide_and_seek_engine.env_wrapper import SARBatchedGridEnv

env = SARBatchedGridEnv(
    num_envs=8,
    map_png="test_level/level.png",
    tiles_json="test_level/tiles.json",
    agents_json="test_level/agents.json",
    survivors_json="test_level/survivors.json",
    map_size=32,
    seed=42,
)

obs, info = env.reset()
actions = env.action_space.sample()
obs, rewards, terminated, truncated, info = env.step(actions)
state = env.state()  # global CTDE state

Observation Space (Local Actor Input)

obs is a dictionary:

• obs["spatial"]: shape [Env, Agent, C_local, H, W]
  • channels include terrain+altitude, local survivor layer, local obs mask, local agent layers
• obs["internal"]: shape [Env, Agent, 6]
  • [deploy_remaining, stuck, view_range, battery, y, x]

State Space (Central Critic Input)

env.state() returns:

• state["spatial"]: shape [Env, C_global, H, W]
• state["internal"]: flattened agent+survivor internal vectors

Action Space (Hybrid)

Per agent action:

• movement: 2D vector in [-1, 1]
• radio: discrete channel 0..3
  • 0 = no transmit
  • 1,2,3 = transmit channel (merged into shared local knowledge)

Tensor shape for partial obs mode stepping batched env:

• [num_envs, n_agents, 3] (dy, dx, radio_channel)

Rendering

• Global view: env.render(env_idx=0)
  • undiscovered tiles are drawn at half RGB brightness
  • saved survivors are white
• Agent POV: env.render_pov(agent_idx=0, env_idx=0)
  • allies and survivors shown using last-known positions
  • knowledge updates when locally seen or shared by radio

Print radio events from current frame:

env.radio_render()

PettingZoo Parallel API

from hide_and_seek_engine.env_wrapper import SARParallelPettingZooEnv

pz_env = SARParallelPettingZooEnv(
    map_png="test_level/level.png",
    tiles_json="test_level/tiles.json",
    agents_json="test_level/agents.json",
    survivors_json="test_level/survivors.json",
)

obs, infos = pz_env.reset()
actions = {
    agent: {"move": [0.0, 1.0], "radio": 1}
    for agent in pz_env.agents
}
obs, rewards, terminations, truncations, infos = pz_env.step(actions)

Test & Benchmark Suite

Run unit checks + 10k-step stress tests + FPS measurements + renderer smoke test:

python env_spec.py --steps 10000 --envs 1 2 4 8

Skip renderer test:

python env_spec.py --steps 10000 --envs 1 2 4 8 --skip-render

Human Data Recorder

Collect SARSA tuples from one human-controlled random agent each episode:

python human_runner.py

To record a visual replay of the session:

python human_runner.py --record

Controls:

• movement: W, A, S, D
• radio: 1, 2, 3

After each episode, enter a save name. Data is written to saved_human_behavior/<name>/, containing:

• .npy files for all observation/state/action/reward buffers.
• replay.gif (if --record was used).