PAOP123/robochunk-sim

RoboChunk simulated data and results

Independent re-implementation of H-Net-style dynamic chunking for robot trajectories. Not affiliated with the H-Net authors (Sukjun Hwang, Brandon Wang, Albert Gu) or Cartesia.

Simulated pick-and-place trajectories with ground-truth motion-phase labels, generated for RoboChunk (https://github.com/NapoleonPK/robochunk-study), plus the study's result files. All data here was generated by the study's own scripted expert; no third-party data is included. Checkpoints: https://huggingface.co/PAOP123/robochunk

Results

test	criterion	key numbers (from runs/results_*.json)	verdict
H1	learned router − every baseline ≥ 0.10 boundary F1 (±3), CI excludes 0	learned 0.358 vs Fixed 0.594; difference −0.236 [−0.253, −0.219]	fail
H2	learned bits/step < Fixed at ≥ 2 of 3 N, compression matched	N=5: -13.165 vs -12.650; N=10: -13.150 vs -12.825; N=20: -13.208 vs -13.009 (compression 15.5 < 0.8 N, counts as fail)	pass
H3	≥ 2× chance near gripper events and p < 0.01, on both datasets	SO-100 1.89× (p = 0.001); Koch 1.77× (p = 0.001)	significant but below threshold
H4	Spearman ρ(novelty, gain) ≥ 0.6 over 11 clean candidates	ρ = 0.782 (p = 0.0045); 95% CI [0.25, 0.95] (post hoc); corrupted data scored most novel in 4 of 4	pass
H1-v2	refractory router − every baseline ≥ 0.10 raw F1 (fresh data)	vs Fixed +0.042 [+0.023, +0.061]	fail (predicted)
H1b-v2	chance-corrected (circular shift) difference > 0 vs every baseline, CI excludes 0	vs Fixed +0.103 [+0.084, +0.122]; smallest (vs Velocity minima) +0.055 [+0.038, +0.072]	pass
H3-v2	refractory router, ≥ 2× chance near gripper events, SO-100 and Koch	SO-100 1.53× (p = 0.001); Koch 1.52× (p = 0.001)	not supported
H3b-v2	same rule on a fresh dataset (SO-101, 5 test episodes)	1.74× (p = 0.004)	fail

True motion phases (bands) and the boundaries each rule places on six simulated test episodes, N = 10.

Full report with all numbers, caveats and deviations: REPORT.md in the GitHub repository.

Splits

FetchPickAndPlace-v4 (gymnasium-robotics 1.4.2, MuJoCo 3.3.7), 25 Hz. The viewer shows the test split (H1–H2) and the fresh split (generated after the v2 pre-registration, used by H1-v2/H1b-v2) as Parquet, one row per step. All splits are also in sim/*.npz (read with robochunk.data_sim.load; regenerate with python scripts/make_sim_data.py).

split	episodes	steps	Parquet	seeds
train	2400	189155	npz only	0–2399
val	300	23759	npz only	100000–100299
test	300	24043	yes	200000–200299
fresh	300	23427	yes	700000–700299 (v2)
pilot	200	16179	npz only	900000–900199

Part B (partb/*.npz): reference set R, R_heldout, R_val, clean candidates C1–C11, corrupted candidates X1–X4 (what each one is: DECISIONS.md in the GitHub repository).

Part B file	episodes
C1	200
C2	200
C3	200
C4	200
C5	200
C6	200
C7	200
C8	200
C9	200
C10	200
C11	200
R	500
R_val	50
R_heldout	100
X1	200
X2	200
X3	200
X4	200

results/: every result file the report is generated from, and the per-episode policy successes behind H4 (partb_policy*/).

Fields (Parquet)

field	meaning
episode_index	episode number within the split (0-based)
episode_seed	seed that regenerates the episode exactly
step	time step within the episode (25 Hz, at most 150 steps)
phase, phase_name	ground-truth motion phase (table below)
phase_start	true if a new phase starts at this step: the boundaries scored by H1 (step 0 excluded)
contact	a finger touches the object
success	object within 5 cm of the goal at this step
episode_retried, episode_truncated	the episode contains a retry / hit the 150-step limit
obs_grip_pos_x … obs_finger_vel_1	the 25 FetchPickAndPlace observation channels: gripper position, object position, object relative to gripper, finger positions, object rotation, object linear/angular velocity, gripper velocity, finger velocities (simulator units)
action_dx … action_gripper	the expert's action: end-effector displacement command (x, y, z) and gripper command (positive opens, negative closes), with Gaussian noise std 0.05
goal_x … goal_z	goal position for the object (constant within an episode)

Phases

id	phase	what the scripted expert does
0	approach	move the open gripper to 10 cm above the object
1	descend	lower the open gripper to the object (on planned retries, offset 4.5–6 cm along x so the grasp misses)
2	close	close the fingers in place
3	lift	raise the gripper 8 cm with the fingers closed
4	transport	carry the object to 1 cm above the goal
5	open	open the fingers in place
6	retreat	raise the open gripper 10 cm

The phase label at a step is the most recently started phase. Each phase follows a minimum-jerk profile with a random speed factor (0.5–1.5), one pause of 0–5 steps, and may start 0–3 steps before the previous one ends. If the object has not risen with the gripper when transport is due to start, the expert opens and repeats approach → retreat (a retry; 10% of episodes plan a missed first grasp).

from datasets import load_dataset
ds = load_dataset("PAOP123/robochunk-sim", split="test")
ep = ds.filter(lambda r: r["episode_index"] == 0)
print(ep["phase_name"][:20])

Pre-registration hashes

SHA-256 of the pre-registration files at each tag and now (shasum -a 256 PREREGISTRATION*.md). PREREGISTRATION.md changed once, by Amendment 1, before any main result; neither file has changed since its last tag.

file	SHA-256	tag (commit)
PREREGISTRATION.md	e78668239e7401270c572d3feaef8d62cb2210e9d582ae640d525c94b52d8850	prereg (7ffc024)
PREREGISTRATION.md	628c6ad61a63b9ee08a84e2d4bf71d6ff3562c84c463cb87767c8c012ec86095	prereg-amend1 (fb5e2ad)
PREREGISTRATION.md	628c6ad61a63b9ee08a84e2d4bf71d6ff3562c84c463cb87767c8c012ec86095	prereg-v2 (9e99328)
PREREGISTRATION_v2.md	fa2700491bb7ca95826a0f81b868b538ebae54b47ba37d609a6e18ebd860c6b7	prereg-v2 (9e99328)
PREREGISTRATION.md	628c6ad61a63b9ee08a84e2d4bf71d6ff3562c84c463cb87767c8c012ec86095	current file
PREREGISTRATION_v2.md	fa2700491bb7ca95826a0f81b868b538ebae54b47ba37d609a6e18ebd860c6b7	current file

Timeline evidence

Timestamps recorded by third parties (UTC). Source details: docs/timeline_evidence.json in https://github.com/NapoleonPK/robochunk-study.

UTC	source	what it shows
2026-10-29 13:08:38	GitHub push event (earlier private repo)	push to main ending at old commit 1f1faa0 (= 7ffc024, tag prereg)
2026-10-29 13:08:51	Gmail, sent check-in email	check-in 1+2; its text quotes 'Pre-registration committed as 1f1faa0 (tag "prereg")'
2026-10-29 18:29:16	Gmail, sent check-in email	'Amendment 1 is committed and pushed (tag prereg-amend1). No main result existed before it.' (tag name, no hash)
2026-10-29 18:35:33	GitHub push event (earlier private repo)	push whose previous head was old commit f61d0f3 (= fb5e2ad, tag prereg-amend1), so that commit was on GitHub by then
2026-10-30 08:27:17	Gmail, sent check-in email	v1 results H1-H3 reported (v2 was designed after this, as PREREGISTRATION_v2.md states)
2026-10-30 08:42:19	Gmail, sent email	draft PREREGISTRATION_v2.md rev. 1 attached, for approval
2026-10-30 08:53:41	Gmail, sent email	draft PREREGISTRATION_v2.md rev. 2 attached, for approval (the approved rev. 3 was not emailed)
2026-10-30 09:42:58	GitHub (public repo created)	NapoleonPK/robochunk-study created with the rewritten history, including tag prereg-v2 (9e99328)
2026-10-30 10:39:31	GitHub Actions (public)	first CI run, on commit 53c390d, which has 9e99328 (tag prereg-v2) as an ancestor; the v2 training runs started after this
2026-10-30 13:42:56	GitHub Actions (public)	CI run on commit 9e1347d, which adds runs/results_v2.json (v2 analysis, single run)
2026-10-30 13:45:55	Gmail, sent check-in email	v2 results reported
2026-10-31 07:12:27	Hugging Face commit	dataset PAOP123/robochunk-sim, first upload (commit f509bea0ce)
2026-10-31 07:13:24	Hugging Face commit	model PAOP123/robochunk, first upload (commits be8a09a8f0, 9f7f28c5ac at 07:13:29Z)

Self-reported (git and local run logs, not independent): 2026-10-29 13:08:32 git commit 7ffc024 (old 1f1faa0), tag prereg; 2026-10-29 18:28:57 git commit fb5e2ad (old f61d0f3), tag prereg-amend1; 2026-10-29 18:29:02 first v1 main run started (runs/main_queue.log; no run had finished before the amendment); 2026-10-30 08:57:49 git commit 9e99328 (old d9b86c7), tag prereg-v2; 2026-10-30 08:58:28 git commit ac72fec: fresh simulated test split generated after the tag, SHA-256 recorded; 2026-10-30 11:14:00 v2 training runs started (runs/after_main.log, POLICIES_DONE 12:14 local time); 2026-10-30 13:40:00 v2 runs finished (V2_DONE 14:40 local time); analyze_v2.py run once afterwards.

History rewrite and commit map. Before the repository was made public (2026-10-30), its git history was rewritten once with git filter-repo --replace-text, only to replace the author's personal e-mail address in docs/BRIEF.md. File contents, dates, messages and tags are otherwise unchanged, so the pre-registration files have the same SHA-256 in the old and the new commits (checked for all three tags), but every commit hash changed. Emails, the earlier private repository and run git.txt files cite the old hashes; docs/HISTORY.md and docs/commit-map.txt map them to the new ones (for example 1f1faa0 → 7ffc024, tag prereg). Gmail: message 'date' from the Gmail API for messages sent from the author's account to the same address (verifiable by the account owner; originals with headers available on request). GitHub events: repository events API of the earlier private repository NapoleonPK/robochunk (visible to its owner while it exists). GitHub Actions and Hugging Face: public.

Credits

The model follows H-Net (Sukjun Hwang, Brandon Wang, Albert Gu, Dynamic Chunking for End-to-End Hierarchical Sequence Modeling, arXiv:2507.07955) and Cartesia's posts on dynamic chunking. Simulation: Gymnasium-Robotics and MuJoCo. Author: Téo Tran Van Lieu. Code written with Claude (Anthropic) as a coding agent.