mettagrid 0.26.20

A fast grid-based open-ended MARL environment

MettaGrid Environment

MettaGrid is a multi-agent gridworld environment for studying the emergence of cooperation and social behaviors in reinforcement learning agents. The environment features a variety of objects and actions that agents can interact with to manage resources, engage in combat, and optimize their rewards.

Requirements

• Bazel 9.0.0 or newer (the project uses Bzlmod and modern Bazel features)
• Python 3.12
• C++ compiler with C++20 support

Overview

In MettaGrid, agents navigate a gridworld and interact with various objects to manage resources, engage in combat, and cooperate with other agents. The key dynamics include:

• Resource Management: Agents carry an inventory of typed resources with configurable capacity limits. Resources can be gathered from objects, consumed by actions, and stolen through combat.
• Combat: Agents can attack other agents by moving into them when their vibe matches an attack configuration. Successful attacks freeze the target and allow the attacker to steal resources. Targets can defend using armor and defense resources.
• Vibes: Each agent has a vibe that determines how they interact with other agents on contact. Vibes affect which attacks trigger, defense bonuses, and are visible in observations.

The environment is highly configurable, allowing for experimentation with different world layouts, object placements, game mechanics, and agent capabilities.

Objects

Agent

The Agent object represents an individual agent in the environment. Agents can move, attack, change their vibe, and interact with other objects. Each agent has an inventory, a vibe, and a frozen state that govern its abilities and interactions.

Wall

The Wall object acts as an impassable barrier in the environment, restricting agent movement.

Custom Objects

Object types beyond Agent and Wall are defined entirely through configuration. Any object can have an inventory, on-use handlers that trigger when an agent moves into them, area-of-effect behaviors that apply to nearby objects each tick, and territory controls.

Actions

Move

The move action allows agents to move to an adjacent cell in the gridworld. Movement supports up to 8 directions (4 cardinal and 4 diagonal), configured via allowed_directions. Moving into different targets triggers different behaviors:

• Moving into an empty cell moves the agent normally.
• Moving into an object with an on-use handler triggers that handler (e.g., gathering resources from a generator).
• Moving into a frozen agent swaps positions with them.
• Moving into an agent whose vibe matches an attack configuration triggers the attack.
• Otherwise, the movement fails.

Attack

Attacks are not standalone actions. They trigger when a move lands on an agent whose vibe matches an attack handler configuration. The attack system uses weapon, armor, and defense resource calculations:

• Weapon power is computed from the attacker's inventory, weighted by configured amounts.
• Armor power is computed from the target's inventory, with optional vibe-based bonuses.
• If the target has sufficient defense resources, the attack is blocked and those resources are consumed.
• On a successful attack, the target is frozen for a configured duration and inventory loot is transferred to the attacker.

Change Vibe

The change_vibe action sets the agent's current vibe. There is one action variant per configured vibe, named after the vibe (e.g., change_vibe_default, change_vibe_junction). Changing vibe always succeeds.

Noop

The noop action is a no-operation. It always succeeds with no side effects.

Handler System

Game logic in MettaGrid is driven by an event-based handler architecture rather than hardcoded per-object behavior. This makes the environment highly composable and configurable.

A handler pairs a chain of filters with a chain of mutations. All filters must pass for the mutations to execute. Handlers are used for on-use interactions (when an agent moves into an object) and area-of-effect behaviors (applied to nearby objects each tick).

A multi-handler dispatches to multiple handlers with either first-match or apply-all semantics.

Filters gate whether a handler's mutations execute. Available filter types include vibe matching, resource thresholds, tag membership, shared tags between actor and target, game value thresholds, distance checks, and boolean combinators (negation and disjunction).

Mutations modify game state when a handler fires. Available mutation types include resource deltas, resource transfers between entities, freezing agents, clearing inventories, weapon-vs-armor attack calculations, stat logging, tag modifications, and inventory queries across multiple objects.

Events fire at specific timesteps, applying mutations to objects matching a query. An event scheduler efficiently dispatches these with minimal overhead when no events are due.

Configuration

The MettaGrid environment is highly configurable through Pydantic-based configuration classes. The top-level MettaGridConfig contains:

• Game rules and episode length
• Action definitions (move, attack, change vibe, noop) with per-action resource requirements
• Agent properties including group, freeze duration, initial inventory, rewards, and per-tick handlers
• Object type definitions with on-use handlers, area-of-effect configs, territory controls, and inventory
• Handler and event definitions with filter and mutation chains
• Observation feature configuration
• Map generation settings

Map Generation

MettaGrid includes a procedural map generation system. The MapBuilder base class has implementations for ASCII-based maps and random generation. The MapGen system provides advanced scene-based procedural generation with composable scene types including biomes (forest, desert, caves, city), shape generators (BSP, maze, spiral, wave function collapse), and transformations (mirror, rotate, copy). Pre-built map configurations are available in configs/maps/.

Environment Architecture

Core Simulation

The Simulation class provides the core simulation for running MettaGrid episodes. It offers direct access to the simulation without an environment API — use it when you need fine-grained control over simulation steps, agent actions, and state inspection. The Simulator class is a factory that creates Simulation instances with map caching and event handler support.

Environment Adapters

MettaGridPufferEnv is the primary PufferLib-compatible environment used by the Metta training system. It provides the standard reset()/step() API with stats collection and supervisor policy support.

MettaGridPettingZooEnv is a PettingZoo ParallelEnv adapter for multi-agent research with standard dict-based observation and action interfaces.

Visualization

Mettascope is a Nim-based GUI viewer for simulation replay and real-time visualization, built as part of the package. Miniscope provides a terminal-based renderer with symbol-based map display.

Compatibility Testing Demos

These demos ensure external framework adapters remain functional as the core environment evolves:

# Verify PettingZoo compatibility
python -m mettagrid.demos.demo_train_pettingzoo

# Verify PufferLib compatibility
python -m mettagrid.demos.demo_train_puffer

The demos serve as regression tests to catch compatibility issues during core development, ensuring external users can continue using their preferred frameworks.

Building and Testing

For local development, refer to the top-level README.md in this repository.

Bazel

By default, uv sync will run the Bazel build automatically via the custom build backend. If you need to run C++ tests and benchmarks directly, you'll need to invoke bazel directly.

Build C++ tests and benchmarks in debug mode:

# Build with debug flags
bazel build --config=dbg //:mettagrid_c
# Run all tests
bazel test //...

For benchmarks you might prefer to use the optimized build:

# Build with optimizations
bazel build --config=opt //:mettagrid_c

For a single-core benchmark of MettaGrid performance (triggers a rebuild on first run):

bash benchmarks/perf/run.sh                    # toy config (default)
bash benchmarks/perf/run.sh --config arena      # production training config

Debugging C++ Code

MettaGrid is written in C++ with Python bindings via pybind11. You can debug C++ code directly in VSCode/Cursor by setting breakpoints in the C++ source files.

Prerequisites

1. VSCode Extension: Install the Python C++ Debugger extension (pythoncpp)
2. Debug Build: Always build with DEBUG=1 to enable debug symbols and dSYM generation

Setup

The repository includes pre-configured launch configurations in .vscode/launch.json:

• MettaGrid Demo and other pythoncpp configurations - Combined Python + C++ debugging session for the demo script (requires the pythoncpp extension)
• _C++ Attach - Attach C++ debugger to any running Python process (shared by all configurations but can be ran manually).

Quick Start

1. Build with debug symbols:

   • Clean everything up

     cd packages/mettagrid # (from root of the repository)
     bazel clean --expunge
     

   • Rebuild with debug flags

     bazel build --config=dbg //:mettagrid_c
     

   • Or Reinstall with DEBUG=1 to trigger dSYM generation

     cd ../..
     export DEBUG=1
     uv sync --reinstall-package mettagrid
     

2. Set breakpoints in both Python and C++ files (e.g., packages/mettagrid/cpp/bindings/mettagrid_c.cpp, packages/mettagrid/demos/demo_train_pettingzoo.py)

3. Launch debugger using the "MettaGrid Demo" or any other pythoncpp configuration from the VSCode Run panel.

4. Alternatively, you can use the "_C++ Attach" configuration to attach the debugger to any running Python process. It will ask you to select a process - type "metta" or "python" to filter the list.

Testing C++ Debugging

To verify that C++ breakpoints are working correctly, use a simple test that calls from Python into C++:

Quick Test Method

1. Add a test call to any Python entrypoint that uses mettagrid:

   def test_cpp_debugging() -> None:
       """Test function to trigger C++ code for debugging."""
       try:
           from mettagrid.mettagrid_c import PackedCoordinate
   
           # Call a simple C++ function
           packed = PackedCoordinate.pack(5, 10)
           print(f"C++ test: PackedCoordinate.pack(5, 10) = {packed}")
   
           # Unpack it back
           r, c = PackedCoordinate.unpack(packed)
           print(f"C++ test: PackedCoordinate.unpack({packed}) = ({r}, {c})")
       except Exception as e:
           print(f"C++ debugging test failed: {e}")
   
   # Call at module level or early in your script
   test_cpp_debugging()
   

2. Set a C++ breakpoint in the corresponding C++ implementation:

   • Open packages/mettagrid/cpp/include/mettagrid/systems/packed_coordinate.hpp
   • Find the pack() or unpack() function implementation
   • Set a breakpoint inside the function body (e.g., on the return statement)

3. Launch your debug configuration (e.g., "MettaGrid Demo" or any pythoncpp configuration)

4. Verify the breakpoint hits when the Python code calls PackedCoordinate.pack()

Where to Add the Test

Add the test call early in any Python entrypoint that uses mettagrid:

• Demo scripts (e.g., packages/mettagrid/demos/demo_train_*.py)
• CLI entrypoints (e.g., packages/cogames/src/cogames/main.py)
• Tool runners (e.g., common/src/metta/common/tool/run_tool.py)
• Training scripts (e.g., metta/tools/train.py)

Note: This test is only for verifying your debugging setup. Remove it before committing.

Configuration Files

• .bazelrc - Defines the --config=dbg build mode with debug flags (-g, -O0, --apple_generate_dsym)
• .vscode/launch.json - Contains launch configurations for combined Python/C++ debugging

Important Notes

• Always use DEBUG=1: Without this environment variable, dSYM files won't be generated and C++ breakpoints won't work.
• Source maps: The launch config includes source maps to correctly locate C++ files in the packages/mettagrid's workspace.