armteam/hapticwam-evidence

HapticWAM — per-take evidence behind the paper's tables

HapticWAM: Distilling Imagined Touch into a World-Action Model without Inference-Time Tactile Sensing — paper arXiv:2609.23888, submitted to ICRA 2027. Code: github.com/Advanced-Robotic-Manipulation/HapticWAM · all repos: HapticWAM — ICRA 2027.

This dataset is the scored evidence a reader needs to recompute the paper's result tables: the per-trial outcome CSVs for the 200 rig trials, the four offline-probe JSONs behind §IV-F, and the plotting script with its rendered figures. Nothing else — no raw episodes, no video, no source snapshots. The raw episodes those scores were computed from are armteam/hapticwam-rig-episodes.

Every file is a copy of the corresponding file in the project's result archive, with one documented exception noted under Provenance.

The paper-to-artifact map, including the exact commands that re-derive each number, is docs/reproducibility.md in the code repository. This card is the short form of it.

What maps to what

File	Bytes	Feeds
rig_0916/per_take_final.csv	47,339	Tables II and III, waffles and Carton — 160 rows, one per rig trial: the operator verdict, the sensor-derived class, the pinch-force peaks, stop reason and replan count
rig_0916/per_take_egg.csv	12,183	Tables II and III, Egg — the same 40 rows for the egg task
rig_0916/summary_by_arm.csv	1,319	the per-arm rollup for waffles and Carton that Tables II and III aggregate to
rig_0916/probe/student_none.json	332,072	§IV-F, intact student — summary.endpoint_err_mm = 22.309 mm (paper: 22.31 mm); the zero_endpoint_err_mm column of rows gives the 28.2 mm no-motion floor
rig_0916/probe/student_contact_zero.json	332,739	§IV-F, contact frames clamped to zero — 28.311 mm (paper: 28.31 mm)
rig_0916/probe/student_prev_cpk.json	332,015	§IV-F, ACC input zeroed — 22.348 mm (paper: 22.35 mm)
rig_0916/probe/student_contact_gt.json	335,777	the fourth probe condition, contact pinned to the ground-truth package — 24.827 mm, not reported in the paper
rig_0916/figures/make_figures_0916.py	12,967	the plotting script; it reads the two per-take CSVs, so its output is an independent check on Tables II and III
rig_0916/figures/fig_rig_outcomes.{png,pdf}	216,981 / 49,137	rendered outcome plot
rig_0916/figures/fig_rig_pinch.{png,pdf}	195,197 / 29,757	rendered pinch-force plot
rig_0916/moves_manifest.csv	7,854	the 81 rig takes recorded on the same day that are not among the 200 scored trials — three arm labels the paper does not report, plus one aborted take. It explains why the episode repository holds more episodes than the tables score

No figure in the published v1 plots the rig numbers; the two figures here are archived as a second path to the same tables, not as paper figures.

Fetching

hf download armteam/hapticwam-evidence --repo-type dataset --local-dir .

Or a single file:

https://huggingface.co/datasets/armteam/hapticwam-evidence/resolve/main/rig_0916/per_take_final.csv

Re-deriving Tables II and III

python - <<'PY'
import csv, statistics as st
B = "rig_0916/"
ARMS = [("v6_simft2k","Teacher"), ("stu_simft_001000","Student"), ("pi05","pi0.5"), ("dp","DP")]
rows = []
for f, fixed in (("per_take_final.csv", None), ("per_take_egg.csv", "Egg")):
    for r in csv.DictReader(open(B + f)):
        low = r["episode"].lower()
        r["task"] = fixed or ("Carton" if "carton" in low else "waffles")
        r["op"] = str(r["operator_placed"]).lower() == "true" or r["verdict"] == "crushed"
        r["sensor_placed"] = r["class"] in ("placed_clean", "placed_crushed")
        rows.append(r)

print("TABLE II  operator placement verdicts")
for a, name in ARMS:
    cells, tot, hit = [], 0, 0
    for t in ("waffles", "Carton", "Egg"):
        sel = [r for r in rows if r["arm"] == a and r["task"] == t]
        k = sum(1 for r in sel if r["op"]); cells.append(f"{100*k/len(sel):3.0f}% ({k:2d}/{len(sel)})")
        tot += len(sel); hit += k
    print(f"  {name:8} " + "   ".join(cells) + f"   pooled {hit}/{tot} = {100*hit/tot:.0f}%")
PY

The full script, including Table III and the paired sign test, is in docs/reproducibility.md §3.

MANIFEST.tsv

MANIFEST.tsv lists every file with four columns:

column	meaning
path	path inside this dataset repository
bytes	file size
sha256	SHA-256 of the file contents
git_blob_sha1	the git blob id of the file (git hash-object <file> reproduces it)

Provenance

The measurements, verdicts, classes and counts in these files are exactly those the paper reports; they have not been recomputed, rounded or re-scored for this release.

One documented edit: in per_take_final.csv, three rows carried a person's name in the free-text notes column, recording whose call a late verdict was entered on. The name and the timezone suffix were removed from those three strings; the verdicts, times and every other field are unchanged, and re-deriving Tables II and III from the edited file reproduces the published numbers exactly. No other file was modified.

Part of the HapticWAM release

Ten repos on the hub, gathered in the HapticWAM — ICRA 2027 collection.

Repo	Kind	Holds
armteam/hapticwam-teacher	model	the tactile-input teacher. Deployed checkpoint teacher_v6_simft/teacher_002000.pt; also holds the Cosmos prompt cache text_embeddings.pt
armteam/hapticwam-student	model	the distilled pad-free student, the model that runs on the rig. Deployed checkpoint hid_simft/student_001000.pt
armteam/hapticwam-baselines	model	the pi0.5, Diffusion Policy and X-VLA baselines at the deployed steps
armteam/hapticwam-ablations	model	every training arm that is not deployed, and the complete evaluation sweeps
armteam/hapticwam-teleop-dataset	dataset	the training corpus — 1,115 teleoperated episodes, packed per task
armteam/hapticwam-teleop-raw	dataset	the same teleoperation as loose, as-recorded sessions (provenance)
armteam/hapticwam-sim-episodes	dataset	Isaac Sim expert episodes, used for the sim fine-tune
armteam/hapticwam-rig-episodes	dataset	the closed-loop rig takes the reported numbers are computed from
armteam/hapticwam-rollouts	dataset	policy-driven rollouts — the DAgger rounds and the deploy days
armteam/hapticwam-evidence ← you are here	dataset	per-take evidence behind the paper's tables — scored CSVs, probe JSONs, figures

Code, training and deployment scripts: github.com/Advanced-Robotic-Manipulation/HapticWAM.

Licence

• rig_0916/figures/make_figures_0916.py: Apache-2.0, the licence of the HapticWAM code repository.
• Everything else (the CSVs, the probe JSONs and the rendered figures): CC-BY-4.0.