ReMDM-MiniHack

Generative Planning Agent for MiniHack navigation using Re-Masked Discrete Diffusion (ReMDM).

The agent uses Masked Discrete Diffusion to iteratively generate action sequences for dungeon navigation. Instead of predicting the next action autoregressively, the model generates entire 64-step trajectories by progressively unmasking action tokens.

Code

GitHub: piotrwilam/ReMDM-MiniHack-Project

Models

Version	Model	Params	Training	Tag
v017_local_baseline	LocalDiffusionPlanner	7M	Offline BC, 200 demos/env, 30 epochs	—
v017_local_baseline	LocalDiffusionPlanner	7M	Offline BC, 500 demos/env, 60 epochs	v0.17-local-baseline-gold (pending)

Repo Structure

ReMDM-MiniHack/
├── README.md                              # This file
├── v017_local_baseline/
│   ├── inference_weights.pth              # EMA state dict (for evaluation)
│   ├── full_checkpoint.pth                # Full training state (for resuming)
│   ├── config.json                        # Hyperparams + model args
│   └── eval_results.csv                   # Per-environment results
└── datasets/
    └── oracle_demos_v017.pt               # Oracle demonstration dataset

Quick Start

import torch
from huggingface_hub import hf_hub_download

# Download weights
path = hf_hub_download("piotrwilam/ReMDM-MiniHack", "v017_local_baseline/inference_weights.pth")
weights = torch.load(path, map_location="cpu", weights_only=False)

# Load model
from model import LocalDiffusionPlanner
model = LocalDiffusionPlanner(action_dim=12)
model.load_state_dict(weights)
model.eval()

Results: v017 Local Baseline (Offline BC, 200 demos/env, 30 epochs)

Environment	Win%	Avg Steps
Room-Random-5x5	94%	18.3
Room-Random-15x15	54%	130.4
Room-Dark-5x5	90%	25.5
Room-Ultimate-5x5	84%	20.8
Room-Ultimate-15x15	30%	72.1
Corridor-R2	42%	132.1
Corridor-R3	0%	200.0
MazeWalk-9x9	48%	119.0
MazeWalk-15x15	22%	162.3