🥮 mooncakes.io

moongridgym

A lightweight Gym-style environment suite for MoonBit.

moonbit
gym
gridworld
maze
rl
moon add zhaoyuanjuns/moongridgym@0.1.1
Download zip
Version
0.1.1
License
MIT
Repository
Last updated
19 hours ago
Downloads
3
README

#MoonGridGym

MoonGridGym is a lightweight Gym-style environment suite written in MoonBit. It bundles a small set of classic grid environments behind one consistent API:

  • GridWorld
  • CliffWalking
  • Maze
  • FrozenLakeLike
  • RandomMaze
  • EmptyRoom
  • FourRooms

Each environment exposes the same three operations:

  • reset()
  • step(action)
  • render()

The suite also ships with a built-in shortest-path solver and rollout helpers for quick demos and evaluation.

The goal is to provide a reusable, deterministic foundation for simulation, planning, search, offline trajectory collection, and reinforcement-learning experiments—not a one-off visual demo.

#Why this is a good OSC2026 topic

  • It is useful on its own, not just as a demo.
  • It sits in a mature area, so the scope can grow without getting narrow.
  • It includes executable benchmark, validation, replay, and dataset tooling.
  • It has fixed and seeded scenarios so results can be reproduced in CI.
  • It exposes numeric observation encoding for downstream agents.

#Quick Start

Requires MoonBit Toolchain v0.10.3 or later.

To obtain the source and run the project locally:

git clone https://github.com/zhaoyuanjuns/MoonGridGym.git
cd MoonGridGym
moon check
moon test
moon run cmd/main

To use the published module from another MoonBit project after the release is available:

moon add zhaoyuanjuns/moongridgym

The module name above matches the authenticated MoonBit account and the GitHub repository owner: zhaoyuanjuns/moongridgym.

moon check
moon test
moon run cmd/main

For the strict local gate, also run:

moon fmt --check
moon info
moon check --deny-warn
moon test --deny-warn

#API

Create a scenario with one of the convenience constructors:

  • new_grid_world(seed)
  • new_cliff_walking(seed)
  • new_maze(seed)
  • new_frozen_lake_like(seed)
  • new_random_maze(seed)
  • new_empty_room(seed)
  • new_four_rooms(seed)

Then use the same interface everywhere:

let env = @moongridgym.new_grid_world(7)
let obs = env.reset()
println(env.summary())
println(obs.ascii)
let result = env.step(@moongridgym.Action::Right)
println(result.info)
println(result.observation.ascii)

#Actions

  • Up
  • Down
  • Left
  • Right
  • Stay

#Scenarios

  • GridWorld focuses on pathfinding with a simple obstacle layout.
  • CliffWalking mirrors the classic control task with a harsh failure state and a positive terminal reward when the goal is reached safely.
  • Maze is a fixed benchmark maze for deterministic examples.
  • FrozenLakeLike adds slipping behavior and holes.
  • RandomMaze generates a seeded maze for replayable experiments.
  • EmptyRoom is a completely empty 7x7 room, ideal for basic random-walk agents.
  • FourRooms is the classic Sutton's 4-rooms environment for hierarchical RL tests.

#Solver Helpers

  • shortest_path() returns a BFS plan from the current agent position.
  • route_string() formats that plan as a readable action sequence.
  • rollout(actions) runs a batch of actions and summarizes the episode.
  • auto_solve() runs the shortest-path plan when one exists.

#Evaluation and data helpers

  • benchmark(seed, episodes) evaluates planner, greedy, and seeded-random policies across all seven scenarios.
  • validate_all(seed) checks reset determinism, reachability, and solvability.
  • collect_episode(...) returns replayable transitions and to_csv() exports an offline-RL-friendly dataset without external dependencies.
  • encode() provides a numeric board representation and checksum.
  • action_checks() and legal_actions() expose boundary and hazard behavior.
  • replay_check(...) detects hidden-state or random-seed regressions.
  • quality_report, contracts_report, and release_gate produce CI-friendly acceptance evidence.

The command-line example prints a scenario summary, a route, a reference rollout, and the local release gate. The library itself remains dependency free and can be embedded in another MoonBit package.

#Reproducible benchmark

let rows = @moongridgym.benchmark(2026, 5)
let report = @moongridgym.benchmark_report(2026, 5)
let data = @moongridgym.collect_episode(
  @moongridgym.ScenarioKind::RandomMaze,
  2026,
  @moongridgym.PolicyKind::ShortestPath,
  512,
)
println(data.summary())
println(data.to_csv())

The seed is part of every benchmark and dataset record. RandomMaze is generated from the seed, while FrozenLakeLike uses the seed for its slip sequence. A repeated seed and action sequence must produce identical replay results; this is covered by the test suite.

#Project boundaries

In scope: discrete 2-D grid environments, deterministic and seeded stochastic transitions, rendering, shortest-path planning, baseline policy evaluation, trajectory capture, numeric encoding, and validation utilities.

Out of scope: neural-network training, external simulators, network services, and a claim of Gymnasium API binary compatibility. These boundaries keep the package portable while leaving clear extension points for future MoonBit agents.

#Current engineering evidence

  • MoonBit is the primary implementation language.
  • The repository contains more than 3,000 lines of maintained .mbt source and tests, including the environment core, planning, benchmark harness, dataset export, replay checks, and acceptance contracts.
  • Seven scenarios, four policy families, seeded benchmark runs, boundary checks, deterministic replay checks, and 36 automated tests are included.
  • CI runs formatting, package info, warning-denied checks, tests, and the runnable example.

#Repository policy

  • MoonBit is the primary implementation language.
  • The code in this repository is original and written from scratch.
  • No upstream source code is copied into this project.
  • If you later port or reference another project, list the upstream source, license, and scope of reuse here before submission.

#Official Mirrors

#Contribution Rules

  • The default branch is master.
  • The commit history is authored by a single contributor account.
  • No virtual or secondary contributors are included.

#Contest readiness checklist

  • Public repository
  • Clear README
  • MIT license
  • Runnable example
  • Automated tests
  • GitHub Actions CI
  • Mooncakes.io publication before final submission

#Mooncakes.io release checklist

Publishing is intentionally not performed by local development commands. After the repository owner authorizes the external release, verify the default branch, confirm the final GitHub/GitLink mirrors contain the same commit, and then publish the module named zhaoyuanjuns/moongridgym from moon.mod. The acceptance package should retain the published version, package page, and the exact commit used for publication as evidence.

#Notes for submission

The OSC2026 guide requires the repository to be public, readable, and actively maintained. It also expects the acceptance materials to show the repository link, README, CI, tests, and Mooncakes publication.

The repository field in moon.mod is already set to the GitHub project URL. Do not publish to GitHub, GitLink, or Mooncakes.io from an unreviewed local working tree.

#
Action

source
pub(all) enum Action {
  Up
  Down
  Left
  Right
  Stay
}

Actions shared by every environment.

#
ActionCheck

source
pub struct ActionCheck {
  action : String
  legal : Bool
  moves : Bool
  reaches_goal : Bool
  lands_in_hazard : Bool
  reason : String
}

Summary of an action's immediate safety without changing the environment.

#
ActionHistogram

source
pub struct ActionHistogram {
  up : Int
  down : Int
  left : Int
  right : Int
  stay : Int
  total : Int
}

Counted action distribution for a collected trajectory.

#
BenchmarkRow

source
pub struct BenchmarkRow {
  scenario : String
  policy : String
  episodes : Int
  successes : Int
  total_steps : Int
  total_reward : Int
  min_steps : Int
  max_steps : Int
  reachable : Int
  planned_steps : Int
}

Aggregate result for one scenario and policy pair.

#
BoundaryScore

source
pub struct BoundaryScore {
  scenario : String
  tested_actions : Int
  blocked_actions : Int
  hazard_actions : Int
  goal_actions : Int
  stable : Bool
}

Boundary and regression counters used by release automation.

#
ContractResult

source
pub struct ContractResult {
  name : String
  passed : Bool
  detail : String
}

A public contract result for downstream packages and release scripts.

#
EncodedObservation

source
pub struct EncodedObservation {
  width : Int
  height : Int
  agent_x : Int
  agent_y : Int
  goal_x : Int
  goal_y : Int
  step : Int
  done : Bool
  cells : Array[Int]
}

A rectangular observation encoding for agents that prefer numeric input.

#
EncodedObservation::checksum

source
fn EncodedObservation::checksum(self : EncodedObservation) -> Int

#
EpisodeDataset

source
pub struct EpisodeDataset {
  scenario : String
  seed : Int
  transitions : Array[Transition]
  total_reward : Int
  success : Bool
}

An in-memory episode dataset. It intentionally uses plain arrays so it is easy to export, inspect, and feed into another MoonBit package.

#
EpisodeDataset::length

source
fn EpisodeDataset::length(self : EpisodeDataset) -> Int

#
EpisodeDataset::summary

source
fn EpisodeDataset::summary(self : EpisodeDataset) -> String

#
EpisodeDataset::to_csv

source
fn EpisodeDataset::to_csv(self : EpisodeDataset) -> String

#
EpisodeScore

source
pub struct EpisodeScore {
  scenario : String
  policy : String
  seed : Int
  steps : Int
  reward : Int
  success : Bool
  terminated : Bool
  truncated : Bool
  reachable : Int
  planned_steps : Int
}

One measured episode. The structure is deliberately small so callers can store many records without depending on a logging framework.

#
EpisodeStats

source
pub struct EpisodeStats {
  steps : Int
  reward_sum : Int
  terminated : Bool
  truncated : Bool
  final_render : String
  info : String
}

Summary information for a rollout or evaluation run.

#
GridGym

source
pub struct GridGym {
  kind : ScenarioKind
  initial_seed : Int
  seed : Int
  width : Int
  height : Int
  start_x : Int
  start_y : Int
  goal_x : Int
  goal_y : Int
  agent_x : Int
  agent_y : Int
  step_count : Int
  step_limit : Int
  done : Bool
  board : Array[Int]
}

#
GridGym::action_checks

source
fn GridGym::action_checks(self : GridGym) -> Array[ActionCheck]

#
GridGym::auto_solve

source
fn GridGym::auto_solve(self : GridGym) -> EpisodeStats

#
GridGym::encode

source
fn GridGym::encode(self : GridGym) -> EncodedObservation

#
GridGym::legal_actions

source
fn GridGym::legal_actions(self : GridGym) -> Array[Action]

#
GridGym::reachable_cells

source
fn GridGym::reachable_cells(self : GridGym) -> Int

#
GridGym::render

source
fn GridGym::render(self : GridGym) -> String

#
GridGym::reset

source
fn GridGym::reset(self : GridGym) -> Observation

#
GridGym::rollout

source
fn GridGym::rollout(self : GridGym, actions : Array[Action]) -> EpisodeStats

#
GridGym::route_string

source
fn GridGym::route_string(self : GridGym) -> String

#
GridGym::shortest_path

source
fn GridGym::shortest_path(self : GridGym) -> PathPlan

#
GridGym::step

source
fn GridGym::step(self : GridGym, action : Action) -> StepResult

#
GridGym::summary

source
fn GridGym::summary(self : GridGym) -> String

#
Observation

source
pub struct Observation {
  kind : String
  ascii : String
  agent_x : Int
  agent_y : Int
  step_count : Int
  done : Bool
}

#
PathPlan

source
pub struct PathPlan {
  found : Bool
  steps : Int
  actions : Array[Action]
}

A path plan returned by the built-in grid solver.

#
PlanExecution

source
pub struct PlanExecution {
  found : Bool
  planned_steps : Int
  executed_steps : Int
  reached_goal : Bool
  reward : Int
  exact : Bool
}

Compare planner length with an executed trajectory.

#
PolicyComparison

source
pub struct PolicyComparison {
  scenario : String
  seed : Int
  best_policy : String
  best_reward : Int
  planner_reward : Int
  greedy_reward : Int
  random_reward : Int
  stay_reward : Int
  planner_success : Bool
  greedy_success : Bool
  random_success : Bool
  stay_success : Bool
}

Per-policy comparison used to choose a baseline before adding a learner.

#
PolicyKind

source
pub(all) enum PolicyKind {
  ShortestPath
  GreedyGoal
  SeededRandom
  Stay
}

Policy families used by the local benchmark harness.

#
QualityScore

source
pub struct QualityScore {
  scenarios : Int
  valid_scenarios : Int
  solvable_scenarios : Int
  deterministic_scenarios : Int
  total_reachable : Int
  total_planned_steps : Int
  passed : Bool
}

A simple scorecard for release readiness. These checks are intentionally executable so CI and maintainers can reuse the same acceptance evidence.

#
ReplayComparison

source
pub struct ReplayComparison {
  same_observation : Bool
  same_reward : Bool
  same_terminal_flags : Bool
  compared_steps : Int
  mismatch_count : Int
}

Difference between two deterministic runs of the same environment.

#
ScenarioKind

source
pub(all) enum ScenarioKind {
  GridWorld
  CliffWalking
  Maze
  FrozenLakeLike
  RandomMaze
  EmptyRoom
  FourRooms
}

The bundled scenario kinds.

#
ScenarioMetadata

source
pub struct ScenarioMetadata {
  kind : String
  description : String
  width : Int
  height : Int
  step_limit : Int
  supports_stochasticity : Bool
  recommended_use : String
}

Static metadata helps applications build scenario pickers without hard-coding descriptions outside the library.

#
SnapshotBatch

source
pub struct SnapshotBatch {
  seed : Int
  values : Array[String]
  checksum : Int
}

A batch of deterministic snapshots for golden-file style testing.

#
StepResult

source
pub struct StepResult {
  observation : Observation
  reward : Int
  terminated : Bool
  truncated : Bool
  info : String
}

#
TrainingCurve

source
pub struct TrainingCurve {
  scenario : String
  policy : String
  rewards : Array[Int]
  successes : Array[Bool]
  cumulative_reward : Int
}

A compact training curve with reward and success observations per episode.

#
TrainingCurve::report

source
fn TrainingCurve::report(self : TrainingCurve) -> String

#
TrainingCurve::successes

source
fn TrainingCurve::successes(self : TrainingCurve) -> Int

#
Transition

source
pub struct Transition {
  scenario : String
  seed : Int
  index : Int
  x : Int
  y : Int
  action : String
  reward : Int
  next_x : Int
  next_y : Int
  terminated : Bool
  truncated : Bool
  info : String
}

A compact transition record suitable for offline RL and regression data.

#
ValidationReport

source
pub struct ValidationReport {
  scenario : String
  valid : Bool
  reachable : Int
  planned_steps : Int
  checks : Int
  failures : Int
  message : String
}

Validation result used by CI and release checks.

#
action_name

source
fn action_name(action : Action) -> String

#
action_roundtrip_count

source
fn action_roundtrip_count() -> Int

Verify that all action names have a stable, documented representation.

#
action_table

source
fn action_table() -> Array[String]

Return a stable action table for consumers that serialize actions by code.

#
all_contracts_pass

source
fn all_contracts_pass(seed : Int) -> Bool

#
all_planner_regrets

source
fn all_planner_regrets(seed : Int, policy : PolicyKind) -> Array[Int]

#
all_plans_valid

source
fn all_plans_valid(seed : Int) -> Bool

#
available_actions

source
fn available_actions() -> Array[String]

#
available_policies

source
fn available_policies() -> Array[String]

Return a compact list of all supported policy names for clients building a command-line or UI selector.

#
available_scenarios

source
fn available_scenarios() -> Array[String]

#
benchmark

source
fn benchmark(seed : Int, episodes : Int) -> Array[BenchmarkRow]

Evaluate every bundled scenario using the deterministic planner and two intentionally weaker baselines. Seeds make the output reproducible.

#
benchmark_report

source
fn benchmark_report(seed : Int, episodes : Int) -> String

#
boundary_report

source
fn boundary_report(seed : Int) -> String

Exercise all five actions from the initial state and return the outcome counters. This explicitly covers boundary, wall, stay, and hazard paths.

#
boundary_score

source
fn boundary_score(kind : ScenarioKind, seed : Int) -> BoundaryScore

#
collect_episode

source
fn collect_episode(kind : ScenarioKind, seed : Int, policy : PolicyKind, max_steps : Int) -> EpisodeDataset

Collect a replayable trajectory using any action policy.

#
collect_reference_dataset

source
fn collect_reference_dataset(seed : Int) -> Array[EpisodeDataset]

Collect one deterministic planner trajectory for every bundled scenario.

#
compare_policies

source
fn compare_policies(kind : ScenarioKind, seed : Int) -> PolicyComparison

#
contract_dataset

source
fn contract_dataset(kind : ScenarioKind, seed : Int) -> ContractResult

#
contract_render

source
fn contract_render(kind : ScenarioKind, seed : Int) -> ContractResult

#
contract_replay

source
fn contract_replay(kind : ScenarioKind, seed : Int) -> ContractResult

#
contract_reset

source
fn contract_reset(kind : ScenarioKind, seed : Int) -> ContractResult

#
contract_solver

source
fn contract_solver(kind : ScenarioKind, seed : Int) -> ContractResult

#
contract_step_progress

source
fn contract_step_progress(kind : ScenarioKind, seed : Int) -> ContractResult

#
contracts_for

source
fn contracts_for(kind : ScenarioKind, seed : Int) -> Array[ContractResult]

#
contracts_passed

source
fn contracts_passed(kind : ScenarioKind, seed : Int) -> Bool

#
contracts_report

source
fn contracts_report(seed : Int) -> String

#
execute_all_plans

source
fn execute_all_plans(seed : Int) -> Array[PlanExecution]

#
execute_plan

source
fn execute_plan(kind : ScenarioKind, seed : Int) -> PlanExecution

#
histogram

source
fn histogram(dataset : EpisodeDataset) -> ActionHistogram

#
histogram_report

source
fn histogram_report(dataset : EpisodeDataset) -> String

#
mean_reward

source
fn mean_reward(kind : ScenarioKind, policy : PolicyKind, seed : Int, episodes : Int) -> Int

Compute the mean reward using integer hundredths, avoiding floating-point dependencies in small embedded targets.

#
new

source
fn new(kind : ScenarioKind, seed : Int) -> GridGym

#
new_cliff_walking

source
fn new_cliff_walking(seed : Int) -> GridGym

#
new_empty_room

source
fn new_empty_room(seed : Int) -> GridGym

#
new_four_rooms

source
fn new_four_rooms(seed : Int) -> GridGym

#
new_frozen_lake_like

source
fn new_frozen_lake_like(seed : Int) -> GridGym

#
new_grid_world

source
fn new_grid_world(seed : Int) -> GridGym

#
new_maze

source
fn new_maze(seed : Int) -> GridGym

#
new_random_maze

source
fn new_random_maze(seed : Int) -> GridGym

#
normalized_episode_count

source
fn normalized_episode_count(episodes : Int) -> Int

Reject unsafe benchmark arguments while retaining a total, predictable API.

#
normalized_step_limit

source
fn normalized_step_limit(limit : Int) -> Int

#
plan_is_valid

source
fn plan_is_valid(kind : ScenarioKind, seed : Int) -> Bool

Verify a plan remains valid when executed one action at a time.

#
planner_regret

source
fn planner_regret(kind : ScenarioKind, policy : PolicyKind, seed : Int) -> Int

A conservative regret estimate relative to the built-in planner.

#
policy_comparison_matrix

source
fn policy_comparison_matrix(seed : Int) -> Array[PolicyComparison]

Produce a comparison matrix with one row per scenario.

#
policy_comparison_report

source
fn policy_comparison_report(seed : Int) -> String

#
project_status

source
fn project_status(seed : Int) -> String

#
quality_report

source
fn quality_report(seed : Int) -> String

#
quality_score

source
fn quality_score(seed : Int) -> QualityScore

Run structural and behavioral checks across the reference scenario set.

#
release_gate

source
fn release_gate(seed : Int) -> String

A short release gate message consumed by the example program and CI logs.

#
replay_check

source
fn replay_check(kind : ScenarioKind, seed : Int, actions : Array[Action]) -> ReplayComparison

Replay an action sequence twice from the same seed and compare every result. This catches accidental hidden state and stochastic reset bugs.

#
run_policy

source
fn run_policy(env : GridGym, policy : PolicyKind, seed : Int, max_steps : Int) -> EpisodeScore

Run a policy with a hard safety cap. The cap prevents a faulty policy from hanging a benchmark runner even when an environment is misconfigured.

#
scenario_catalog

source
fn scenario_catalog(seed : Int) -> Array[ScenarioMetadata]

#
scenario_catalog_report

source
fn scenario_catalog_report(seed : Int) -> String

#
scenario_name

source
fn scenario_name(kind : ScenarioKind) -> String

#
scenario_table

source
fn scenario_table() -> Array[String]

Return a stable scenario table for metadata and dataset headers.

#
snapshot

source
fn snapshot(kind : ScenarioKind, seed : Int) -> String

A compact benchmark row for a single seed, useful for snapshot tests.

#
snapshot_batch

source
fn snapshot_batch(seed : Int) -> SnapshotBatch

#
snapshot_report

source
fn snapshot_report(seed : Int) -> String

#
success_rate

source
fn success_rate(kind : ScenarioKind, policy : PolicyKind, seed : Int, episodes : Int) -> Int

Measure how often a baseline reaches the target within its safety cap.

#
total_regret

source
fn total_regret(seed : Int, policy : PolicyKind) -> Int

#
train_curve

source
fn train_curve(kind : ScenarioKind, policy : PolicyKind, seed : Int, episodes : Int) -> TrainingCurve

#
validate

source
fn validate(kind : ScenarioKind, seed : Int) -> ValidationReport

Validate invariants that matter before a scenario is used as benchmark data: reset determinism, non-empty reachability, and a solvable goal.

#
validate_all

source
fn validate_all(seed : Int) -> Array[ValidationReport]

#
validation_report

source
fn validation_report(seed : Int) -> String

Source Files

Powered by MoonBit

Site sourceReport issuePackagesBuild queueSkillsStatistics

© 2026 mooncakes.io