Who needs tutorials? This AI learns Minecraft through pure trial and error
An AI agent that builds its knowledge graph by experimenting, just like humans.
ADAM learns Minecraft's technology tree through causal discovery, not pre-existing knowledge
📚 https://arxiv.org/abs/2410.22194
🤖 Original Problem:
Existing Minecraft agents struggle with continuous learning and structured knowledge acquisition. They rely heavily on black-box models or pre-existing knowledge, limiting their interpretability and generalization capabilities.
-----
🔧 Solution in this Paper:
→ ADAM: An embodied causal agent with four key modules:
- Interaction Module: Executes actions and records data
- Causal Model Module: Uses LLM-based and intervention-based causal discovery
- Controller Module: Plans and executes tasks using learned causal graphs
- Perception Module: Uses multimodal LLMs for human-like perception
→ Key Features:
- Builds causal knowledge from scratch without prior knowledge
- Uses interventions to verify causal relationships
- Aligns with human gameplay by avoiding omniscient metadata
-----
💡 Key Insights:
→ Integration of causal methods with embodied agents improves interpretability
→ Intervention-based verification enhances causal discovery accuracy
→ Human-like perception without metadata is possible using multimodal LLMs
→ Bootstrapping process enables continuous knowledge expansion
-----
📊 Results:
→ 2.2x speedup vs SOTA in obtaining diamonds in standard Minecraft
→ 4.6x speedup in obtaining raw iron in modified environments
→ Nearly perfect technology tree construction while others show 30% errors
→ Only agent able to obtain diamonds in modified Minecraft recipes
Share this post