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.

44 searches · 14 tasks · 2,165 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	21 searches
test	the benchmark's own evaluation seeds (40000+) — what a scorer is measured on	23 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"]

Tasks

One row per task: how many searches it contributed, which splits, which scene configs, the seeds, and the benchmark commit whose code built those scenes.

That last column is not bookkeeping. The benchmark's success criteria change over time — put_milktea_on_shelf gained an upright requirement, put_milktea_next_to_laptop a 15° tolerance — so two runs of one task under different commits are scored by different rules and should not be pooled. Runs of a task the change did not touch stay comparable.

task	runs	splits	configs	seeds	benchmark commit
drop_apple_in_bin_ks	6	test, train	kitchens_clean, kitchens_d10, kitchens_d15	40003, 40005, 50000	RoboPRO @ 120b0e0, RoboPRO @ 64840ce
move_pen_to_box	2	test, train	study_clean	40000, 50000	RoboPRO @ 64840ce
move_seal_next_to_box	2	test, train	study_clean	40000, 50000	RoboPRO @ 64840ce
move_seal_onto_table	2	test, train	study_clean	40000, 50000	RoboPRO @ 64840ce
pick_apple_from_bowl_ks	2	test, train	kitchens_clean	40003, 50000	RoboPRO @ 64840ce
pick_bottle_from_fridge	2	test, train	kitchenl_clean	40000, 50000	RoboPRO @ 64840ce
pick_boxdrink_from_basket	2	test, train	kitchenl_clean	40000, 50000	RoboPRO @ 64840ce
put_bottle_in_basket	2	test, train	kitchenl_clean	40001, 50000	RoboPRO @ 64840ce
put_bottle_in_fridge	2	test, train	kitchenl_clean	40000, 50000	RoboPRO @ 64840ce
put_bread_on_board_ks	2	test, train	kitchens_clean	40003, 50000	RoboPRO @ 64840ce
put_milktea_next_to_laptop	7	test, train	office_clean, office_d10, office_d15, office_d6	40000, 50000	RoboPRO @ 64840ce
put_milktea_on_shelf	7	test, train	office_clean, office_d10, office_d15, office_d6	40000, 40002, 50000	RoboPRO @ 64840ce
put_phone_next_to_cube	2	test, train	office_clean	40000, 50000	RoboPRO @ 64840ce
put_phone_on_holder	4	test, train	office_clean, office_d10, office_d6	40000, 50000	RoboPRO @ 64840ce

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	40000–50000
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	2,165	44	11
steps	106,050	44	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")   # the 50000+ collection seeds
test  = load_dataset("mahgoobi/sim-branching-data", "nodes", split="test")    # the benchmark's 40000+ bank

The library's splits are wired to the seed blocks, so split="train" gives the collection seeds and split="test" the benchmark's evaluation bank — the same partition the split column of meta/searches records, which stays the authority if the two ever disagree.

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.