Skip to content

Seren666/openworld-minecraft-agent-lab

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Open-World Minecraft Agent Lab

This is an unofficial student research repository for reading notes, repository audits, lightweight reproductions, and toy demos around open-world Minecraft agents.

Research focus:

planning -> memory -> visual grounding -> action representation -> hierarchical action-space learning

My research interest is how these interfaces connect in embodied decision-making for Minecraft-like environments.

This repository starts from CraftJarvis planning/memory/grounding/action-interface works and is being extended toward OpenHA/CrossAgent, GRPO post-training, and world-model-based agent learning.

Current Focus

  • Long-horizon planning
  • Memory-augmented planning
  • Visual-temporal grounding
  • Vision-language-action representation
  • Hierarchical action-space learning and GRPO-style post-training
  • World-model and diffusion-model reading for future agent learning

For Quick Review

Progress

Paper / System Focus What I did Status Link
DEPS / MC-Planner Planning interface Audited the relevant repository and built a lightweight DEPS-style planning demo for Minecraft tasks. Bounded planning pass complete demo, note
JARVIS-1 Memory interface Audited the official repository, inspected fixed memory through aggregate stats, and built a toy memory-augmented planner. Audit and lightweight demo complete audit, demo, note
ROCKET-1 Visual grounding / mask interface Audited the official repository and built a synthetic visual-temporal context prompting demo with mask scenes. Audit and lightweight demo complete audit, demo, note
JARVIS-VLA Action representation / VLA interface Audited the official repository, built a toy action-interface demo, and ran a tiny GPU action scorer. Audit and lightweight demo complete audit, demo, note
OpenHA / CrossAgent Exploratory action-space taxonomy Analyzed metadata/config-derived task-name records, clarified source scope, added a 100-sample sanity check, and summarized failure-mode observations. Preliminary analysis complete analysis, artifacts
World models / diffusion Predictive simulation for planning and RL Added concise next-reading notes for imagined rollouts, failure prediction, data generation, and safer RL. Conceptual note complete note
MineStudio minimal workflow smoke test Minecraft agent toolchain smoke test Package-level install/import passed; cv2 and Java blockers resolved; simulator/rollout still blocked by torch and DISPLAY/Xvfb runtime stack. Smoke test and blocker-reduction complete report, artifacts
GPU sanity Infrastructure Validated a modern PyTorch CUDA 12.8 environment on the Blackwell GPU. Infrastructure check complete artifacts

What This Repository Demonstrates

  • Repository audits for Minecraft open-world agent papers and codebases.
  • Lightweight reproduction-style toy demos that clarify planning, memory, visual grounding, and action representation ideas.
  • Bounded feasibility checks for heavier official systems.
  • A conceptual bridge from action representation to hierarchical action-space learning, GRPO post-training, and world-model reading.
  • A record of blockers for complete end-to-end reproduction work.

What This Repository Does Not Claim

  • It is not an official implementation from the paper authors.
  • It does not claim complete reproduction of official Minecraft agent systems.
  • It does not claim that official Minecraft evaluations were reproduced.
  • It is not a training repository for large Minecraft agent models.
  • It does not commit model weights, checkpoints, datasets, raw videos, secrets, or external repository code.

Key Takeaways So Far

  • One-shot language planning is fragile in long-horizon Minecraft tasks.
  • Memory can help avoid repeated prerequisite mistakes, such as invalid tool use or missing fuel.
  • Language-only subgoals lose spatial detail when multiple objects or affordances are visible.
  • VLA-style action representation shifts the question from "what should be done" to "which action should be selected next".
  • OpenHA/CrossAgent shift the next question to "which action abstraction should be selected now".
  • World models suggest a later direction for predictive simulation, imagined rollouts, and safer RL.
  • Official systems require substantial dependencies, checkpoints, datasets, and Minecraft rendering; this repo focuses on lightweight, honest preparation.

Research Story

The current project narrative is:

planning -> memory -> visual grounding -> action representation -> hierarchical action-space learning

The main synthesis documents are docs/four_interface_comparison.md and docs/latest_direction_bridge.md.

Project Documents

Repository Layout

papers/       Reading notes for core papers and systems
experiments/  Lightweight reproductions and experiment records
docs/         Research maps, summaries, setup notes, and reproduction-scope records
scripts/      Small helper scripts for setup, demos, and log collection
assets/       Figures, diagrams, and small presentation media

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors