1rahu/mini-coffee-test-trajectories

Mini Coffee-Test Trajectories

200 episodes of "unfamiliar house → find kitchen → make coffee," solved by a privileged BFS oracle from the Mini Coffee-Test harness. I generated this mostly as a byproduct of building the harness itself — once the oracle agent existed, dumping its rollouts to a dataset was nearly free.

Each JSONL record is one full episode:

{
  "seed": 0,
  "house": {"entrance_id": ..., "kitchen_id": ..., "rooms": {...}},
  "trajectory": [
    {"obs": {...}, "action": {"verb": "MOVE", "arg": 3}, "reward": 0.2,
     "done": false, "stages": {"entered_kitchen": true, ...}},
    ...
  ],
  "success": true,
  "num_steps": 5
}

house is the ground-truth room graph — types, objects, connectivity — which is useful if you want to train a memory/mapping module with real supervision, or hold it back and only give an agent the per-step obs. trajectory is just the oracle's shortest-path solve, so treat it as a ceiling to compare against, not something to imitate blindly — it's not RL-trained, it's privileged BFS.

Regenerating or scaling up

python generate_dataset.py --episodes 2000 --rooms 12 --out big.jsonl

Loading it

from datasets import load_dataset
ds = load_dataset("json", data_files="coffee_test_trajectories.jsonl")

One thing worth knowing

This comes out of the symbolic gridworld in 01-mini-coffee-test, not a photorealistic simulator, so there are no images here. It's meant for prototyping navigation/memory logic cheaply before paying the compute cost of a real visual backend (AI2-THOR/Habitat).