[Showcase] Emergent 3-phase curriculum from an intrinsically-motivated agent

[augo-augo]

Hi Crafter team,

Thank you for building this benchmark, it’s been incredible to work with.

I wanted to share something unexpected that emerged while training an intrinsically-motivated agent in Crafter. The agent (DTC - Dual-Timescale Competence) learns without external rewards, using only a competence-based intrinsic signal.

What emerged was a clear three-phase developmental pattern:

Phase 1 (0–80k steps): Skill Mastery
The agent masters basics like wake_up and collect_wood, practicing them heavily.

Phase 2 (80k–140k steps): Boredom Trough
My competence reward (difference between slow/fast novelty EMAs) drops to zero as skills are mastered. The agent gracefully deprecates these skills and stops practicing them—you can see wake_up activity drop off completely.

Phase 3 (140k+ steps): Exploration Cascade
The “boredom” state triggers adaptive exploration (boosted entropy via CognitiveWaveController), leading to a cascade of compositional discoveries:

• collect_sapling
• place_plant
• make_wood_sword
• defeat_zombie

I haven’t seen this pattern documented before, but it might speak to the richness of skill composition in Crafter’s design.

Links:

• Training run (all 22 achievements): https://wandb.ai/bsingleton-art-scgwt/scgwt-crafter
• Code: https://github.com/augo-augo/DTC-agent

Thanks for your work on this!

Augo