Eaglerite
netherite's recipe in a browser tab: an Eaglercraft-style world rewritten in JavaScript, a policy that imitates a scripted player, then GRPO. The task is netherite's headline one: a wooden pickaxe from a fresh world, inside 60 seconds of game time.
Live: https://eaglerite.vercel.app
Why a rewrite
netherite rewrites Minecraft 1.7.10 in C and CUDA, exact against the Java client, because the real game runs one world in real time and RL needs thousands. The same holds for Eaglercraft. So training does not happen inside the Eaglercraft client: the parts of Minecraft 1.8.8 (the version EaglercraftX runs) that the task touches are rewritten in plain JS, and 64 worlds step together in one Web Worker.
| netherite | eaglerite |
|---|
| Game | Minecraft 1.7.10 in C/CUDA, bit-exact vs Java | Minecraft 1.8.8 subset in JS, checked against 1.8's numbers |
| Observation | pixels, or a semantic camera | semantic camera: 9 × 5 rays (block class and distance) |
| Imitation | human speedrun footage | a scripted teacher, then DAgger |
| RL | GRPO, clip 0.2, KL β 0.04, 2 epochs | the same, with Dr. GRPO advantages |
| Policy | 181M parameters, 4 GPUs | 50k parameters, one CPU thread |
| Action compiler | GUI macros, hold_attack with mine-lock, look_target | the same three levers |
The engine (engine.js)
- 64 × 64 × 32 worlds. Terrain from improved Perlin noise; oak (WorldGenTrees) and birch (WorldGenForest) grown with vanilla's canopy rule and draw order, using an exact port of
java.util.Random. Terrain is Minecraft-like, not seed-exact.
- Player movement follows
EntityLivingBase / EntityPlayer 1.8.8: walk 4.317 m/s, sprint 5.612 m/s, jump apex 1.249 blocks (1.8's 0.005 motion cutoff), collision Y then X then Z.
- Mining follows
PlayerControllerMP: the click tick, float32 damage, the 5-tick hit delay, 5× slower in the air, survival reach 4.5. A log by hand takes 61 ticks.
- Dropped items follow
EntityItem: spawn offset and velocity, 10-tick pickup delay, pickup box grown by (1, 0.5, 1).
- Crafting is a macro head: planks, sticks, crafting table, place the table, craft the pickaxe (needs the crosshair on a placed table).
What the policy sees and does
- Observation (309 numbers): 45 rays with block class and distance, the two nearest dropped items, the placed table, the crosshair target and break progress, body state, inventory, which macros would work right now, and the clock.
- Actions, one step = 4 ticks: movement (8 options), look (46: keep the view, or turn onto what one of the 45 rays hits), attack (holds until the block breaks), macro (6, invalid ones masked out).
- Network: MLP 309 → 128 → 64 → 62 logits with tanh, hand-written backprop and Adam (
policy.js).
Training (train.js)
- Teacher: 8 rollouts × 64 worlds of a scripted player that knows the map (Dijkstra over columns, punches through leaves, collects drops, crafts). It gets the pickaxe in 200/200 test worlds, mean 17.8 s.
- Behaviour cloning: 20 epochs on its frames.
- DAgger: 3 rounds. The student drives (the teacher takes the wheel 50%, 25%, 12.5% of the time), the teacher labels every state, 8 epochs each.
- GRPO: 8 groups × 8 worlds, each group sharing a seed and a start. Reward: milestones (logs up to 3, planks, table, placed, sticks) plus 5 for the pickaxe plus 10 × the fraction of the minute left. Advantages are centred within the group without dividing by its std (Dr. GRPO), clipped ratio 0.2, KL(π ‖ π_imitation) with β 0.04, 2 epochs, Adam at 2e-4.
On 256 held-out worlds (node, one run), the imitation policy gets the pickaxe 83% of the time and GRPO takes it to 91–94% within about 20 updates. The whole imitation stage costs about 40 s of CPU.
CPU budget
Training runs in a single Web Worker that works in 14 ms slices and sleeps so that busy ÷ (busy + idle) matches the setting. The views are 128 px, redraw at most 20 times a second and only when something moved, and training pauses when the tab is hidden. Measured for the whole of Chrome on an M-series MacBook Air:
| Setting | CPU |
|---|
| Chill | ~27% of one core |
| Normal (default) | ~50% of one core |
| One core | ~82% of one core |
| Watching at 1× | ~12% of one core |
| Paused | ~3% of one core |
Running it
There is no build step. Serve the folder and open it:
python3 -m http.server 8000
# then http://localhost:8000
engine.js, teacher.js, policy.js and train.js have no DOM dependencies and run in node, e.g. new Trainer().work() in a loop.
| File | What it is |
|---|
engine.js | the game: worlds, physics, mining, items, crafting, observation, reward |
teacher.js | the scripted player used for demonstrations and DAgger labels |
policy.js | the MLP, the action heads, Adam |
train.js | demonstrations → cloning → DAgger → GRPO, as small units of work |
worker.js | the training thread and its CPU budget |
render.js | WebGL2 renderer, procedural textures, chunk meshing |
app.js, index.html | the page: views, controls, charts, side by side, play mode |
Notes
Not affiliated with Mojang, Microsoft, Eaglercraft or netherite. Textures are drawn procedurally in render.js; no game assets are included.
