sim-branching-data

At a state: several action chunks proposed from it, and how each one actually ended. A branch the search dropped was cut off mid-episode, so it is resumed from its own snapshot and carried to a finish — the action nobody executed still gets an answer to would this have worked.

28 searches · 14 tasks · 1,354 nodes, each with its own state and image.

This repo hosts the data. What it means, how it was produced and how to use it live in the code that wrote it: https://gitlab.com/mahgoobi/rewind — see its README for the format, the search strategies, and worked examples of loading a record.

Splits

Both halves live here, told apart by the split column of meta/searches rather than by path. The blocks are disjoint by construction, so a scorer trained on one can be measured on the other without having seen it.

split	seeds	in this repo
train	collection seeds (50000+) — fans to learn from	14 searches
test	the benchmark's own evaluation seeds (40000+) — what a scorer is measured on	14 searches

from rewind.record.hub import index
runs  = index("mahgoobi/sim-branching-data")
train = [r for r in runs if r["split"] == "train"]
test  = [r for r in runs if r["split"] == "test"]

Contents

tasks	drop_apple_in_bin_ks, move_pen_to_box, move_seal_next_to_box, move_seal_onto_table, pick_apple_from_bowl_ks, pick_bottle_from_fridge, pick_boxdrink_from_basket, put_bottle_in_basket, put_bottle_in_fridge, put_bread_on_board_ks, put_milktea_next_to_laptop, put_milktea_on_shelf, put_phone_next_to_cube, put_phone_on_holder
scene seeds	40000, 40001, 40003, 50000
search	branch_once, fan of 10, horizon 6
policy	pi05 — robopro @ 30000
cameras	countertop_camera, right_camera, left_camera
action chunk	50 steps

table	rows	files	columns
nodes	1,354	28	11
steps	66,300	28	6

Outcomes, best to worst: hard_success solved it cleanly, soft_success solved it after a collision, soft_failure missed, hard_failure missed and collided. terminal says whether the episode had ended when the outcome was read — tier is an outcome only where it is true.

Loading it

from datasets import load_dataset
nodes = load_dataset("mahgoobi/sim-branching-data", "nodes", split="train")   # HF split, not ours

split="train" above is Hugging Face's own split name — this dataset has exactly one, and it is unrelated to the split column, which is what separates train from test here.

Every node carries its own state — poses, the robot's command, and one JPEG per camera — so nodes alone answers most questions. steps is what happened between two nodes. There is no video: rewind video builds one from these frames when you want to watch a branch.

Data is partitioned as data/<table>/task=<task>/<search_id>.parquet, so one task is one directory:

from huggingface_hub import snapshot_download
snapshot_download("mahgoobi/sim-branching-data", repo_type="dataset",
                  allow_patterns=["meta/**", "data/*/task=drop_apple_in_bin_ks/*"])

Each run's config is stored verbatim at meta/configs/<search_id>.yml, so any run can be repeated from the record itself.

Full format, and everything else: https://gitlab.com/mahgoobi/rewind.