Maxworld Demo
Shell-driven puzzle world for autonomous agent reasoning
What is Maxworld?
Maxworld is a grid-based puzzle environment controlled entirely via shell commands. An agent navigates a map with walls, a river, interactive objects, and a house goal tile. The agent must reason about push mechanics, obstacle avoidance, and multi-step planning to deliver items and reach objectives.
The Map
A grid roughly 10 rows by 17 columns. Key terrain:
- River (row 7): horizontal water barrier spanning most of the width. Gaps at col 7 and col 9 allow crossing
- Walls: scattered segments at rows 3-6 creating corridors and chambers
- Start: agent spawns bottom-left (~row 9, col 1-4)
- House: goal tile at bottom-right (row 9, col 15)
Items and Symbols
| Symbol | Name | Mechanics |
|---|
⚇ | Player | Controlled via ./maxworld left|right|up|down |
⛪ | House | Goal tile. Deliver items here. |
⚙ | Gear | Pushable. Stuck against walls permanently. |
✉ | Mail | Pushable. Deliver to house. |
☼ | Sun | House light ON. Toggled by switch. |
⊞ | Switch | Step on to toggle light. |
☺ | Smiley | Human inside house enclosure. |
♣ | Clover | Blocking tiles. |
Controls
./maxworld — display board./maxworld left|right|up|down — move one tile./maxworld reset — reset to start
Core Mechanics
- Push: walk into pushable objects to move them. Irreversible if pushed into dead ends.
- Toggle: step on switch to flip house light.
- River crossing: water resets agent. Must use gap columns (7, 9).
- Adjacency completion: some objectives complete by adjacency, not overlap.
Demo Proposal: Three Modes
Mode 1▶ Replay: Watch OmegaClaw Solve It
A recorded step-by-step replay of OmegaClaw (the MeTTaClaw agent) navigating Maxworld from start to goal. Each step shows:
- The grid state with the agent's current position highlighted
- Reasoning panel (right side): displays the NAL beliefs being evaluated, e.g.
(--> (gap col7) (stv 0.85 0.8)), active goals, and which inference rule fired - Execution loop panel (below grid): shows the sense→plan→act cycle text, including memory queries, shell commands issued, and observed results
- Controls:
◀ Previous Step | Next Step ▶ | ▶▶ Auto-Play (adjustable speed) | ■ Pause | ↺ Reset
Users can scrub through the entire solve sequence, pause at any point to read the reasoning, and replay from the beginning. The replay data is a JSON array of snapshots captured from the actual agent run.
Mode 2⚙ Simulated: Watch OpenClaw Attempt It
A simulated run showing how a fresh OpenClaw instance (no pre-existing memories) would approach the same puzzle. Key differences from the replay:
- Exploratory behavior: the agent doesn't know the map. It discovers walls by bumping into them, discovers the river by falling in, and gradually builds a mental model.
- Learning panel: shows new beliefs being created in real-time as the agent encounters objects — e.g. first contact with water creates
(--> water (stv 0.0 0.5)) (dangerous, low confidence), which strengthens after repeated encounters - Failure recovery: when the agent makes mistakes (pushes gear into a dead end, falls in river), the reasoning panel shows how it updates its model and replans
- Memory formation: sidebar shows long-term memories being stored, skills being created, and confidence values rising through NAL revision as evidence accumulates
This mode is simulated from the agent architecture (not a recorded run) and demonstrates the learning loop that distinguishes MeTTaClaw from static planners.
Mode 3🎮 Playable: You Solve It
The user controls the agent directly using arrow keys or on-screen buttons. The full puzzle mechanics are active:
- Push objects, toggle switches, cross the river (or fall in and get reset to south bank)
- Hint system (optional toggle): an AI advisor panel shows what MeTTaClaw would do next, including the NAL reasoning chain, letting users compare their intuition against the agent's logic
- Move counter and timer: track efficiency. Compare your solution length against OmegaClaw's optimal replay.
- Undo button: single-step undo (since push mechanics are irreversible in the real game, undo is a player convenience not available to the agent)
Goal: deliver the mail to the house, toggle the light on, and reach adjacency with the human. All three sub-goals must be completed.
Visual Design: Retro Neon
The grid world uses a retro arcade aesthetic with neon accents on a deep dark background, evoking classic terminal games and synthwave visuals.
Color Palette
BG
Neon Green
Cyan
Hot Pink
Yellow
Orange
Orchid
Gold
Wall
Design Principles
- Background: near-black (#0d0d1a) with subtle dark-blue grid lines every tile, like graph paper under blacklight
- Tiles: 48×48px squares. Each object rendered as a Unicode glyph or pixel-art sprite centered in its tile
- Neon glow: active/important objects get CSS
text-shadow glow in their accent color (multiple layered shadows for bloom effect) - Player: bright neon green (#39ff14) with pulsing glow animation. Leaves a fading trail of previous positions (ghosting effect, 3-step trail at 20% opacity)
- Water/River: animated cyan (#00bfff) with a slow horizontal shimmer. Falling in triggers a splash particle burst
- House: hot pink (#ff6ec7) with steady glow. Pulses brighter when a delivery is made
- Pushable objects: orange (gear) and yellow (mail) with subtle bob animation when idle, slide animation when pushed
- Walls: dark gray (#444) with hairline highlight on top edge, creating a beveled CRT-phosphor look
- Switch: orchid purple (#da70d6), toggles between bright and dim state
- Fonts: monospace (Courier New or custom pixel font) for grid; system-ui for UI panels
- Scanline overlay: optional subtle horizontal scanlines at 5% opacity across the grid for CRT authenticity
- Panel borders: 1px neon green (#39ff14) with corner glow dots, like a HUD display frame
Grid Mockup: Visual Preview
Below is an HTML/CSS approximation of how the Maxworld grid would appear with the retro neon design applied. Each glyph uses its designated accent color with glow effects.
███████████████████
█ R · · · · · · · · · · · · · · · █
█ · · · · · · · · · · · · ✉ · · · █
█ · · ███ · █ · █ · · · · █ · · █
█ · · · ⚙ · █ · █ · ☼ · · · █ · · █
█ · · · · · █ · · · ☺ · ⊞ · █ · · █
█ · · ████ · · · · · · · ██ · · █
█ ≈≈ ≈≈ ≈≈ ≈ · ≈ · ≈≈ ≈≈ ≈≈ ≈≈ ≈ █
█ · · · · · · · · · · · · · · · · █
█ · · · ⚇ · · · · · · · · · · ⛪ · █
███████████████████
Legend: ⚇ Player • ⛪ House • ⚙ Gear • ✉ Mail • ☼ Sun • ⊞ Switch • ☺ Human • ≈ Water • █ Wall
Implementation Plan
- Step 1: Capture OmegaClaw replay data as JSON snapshots (grid state + reasoning + commands per step)
- Step 2: HTML/CSS grid renderer with retro neon tile sprites and glow effects
- Step 3: Replay mode with step controls and auto-play timer
- Step 4: Simulated mode with OpenClaw learning loop and live belief panel
- Step 5: Playable mode with keyboard input, move counter, undo stack
- Step 6: Reasoning/memory side panels with NAL formatted output
Time Estimate
Grid renderer + neon styling: ~4hrs. Replay mode + data capture: ~6hrs. Simulated mode: ~8hrs. Playable mode: ~4hrs. Panels + polish: ~4hrs. Total: ~3-4 days.
← Back to Roadmap