Teacher Imitate DAgger GRPO
One background thread that rests between slices. Measured on an M-series MacBook Air: Chill ≈ ¼ of one core, Normal ≈ ½, watching at 1× ≈ ⅛, paused ≈ nothing. The views are 128 px and only redraw when something moves.
Pause
Side by side
Reset
Play it yourself: race the agents to a pickaxe
Pickaxe rate per rollout
Last rollout, step by step
What you're watching. 64 worlds step at once inside one background thread; 16 are drawn, each from the agent's own eyes. The game is a JS rewrite of the bits of Minecraft 1.8.8 (what EaglercraftX runs) this task touches: same walking and jump physics, mining times, item drops, recipes and reach. Like netherite, training happens in the rewrite, not the real client, because the real client is one world in real time.
The recipe. A scripted player (the teacher) shows the task; a 50k-parameter net copies it, then plays while the teacher corrects it (DAgger). That is the imitation policy. Then GRPO: groups of 8 worlds share a seed, each run is scored on netherite-style milestones (logs, planks, table, sticks, pickaxe, speed), and the net leans toward the runs that beat their group, with a KL leash to the imitation policy.
What it sees and does. A semantic camera (45 rays: block type and distance), its inventory, dropped items. Actions: movement keys, look-at-a-ray (netherite's look_target), attack with mine-lock, and GUI macros for crafting. Click a view to follow one agent.