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.

Renamed from armteam/phantom-episodes on 2026-09-19, when the project's working name PHANTOM became HapticWAM (Haptic World-Action Model). The old id still redirects. The Python package and CLI keep the name phantom, so task keys, checkpoint names and config keys are unchanged.

Tactile manipulation episodes for HapticWAM (Haptic World-Action Model), the tactile world-action model: UR3 + Robotiq 2F-85 + 2x Daimon DM-Tac W2L fingertip sensors + RealSense scene camera, teleoperated via the Echo exoskeleton leader.

Layout

tasks/<task>/ep_<task>_<epoch>_<idx>/ — canonical flattened view: one directory per episode, grouped by task. Use this + manifests/.
manifests/<task>.jsonl, manifests/all.jsonl — one JSON object per episode: success, split (train/val), tactile_contact, duration, peak force, contact area, source session, quality notes, hub path.
manifests/quality_full.csv — full per-episode quality audit.
archive/, collect/ — raw as-recorded session layout (provenance; superset that also contains pre-cleanup debug takes). Directory names are the raw session labels as typed at the rig (carton, *_fail_undergrasp); meta.json is authoritative for task/labels.
archive/20260822_* — batch batch_20260822 (34 sessions / 325 eps), NOT yet in tasks/ or manifests/. Appended at provision time by tools/intake_recovery.py (normalize → place → manifest; all rows go to train, the v4 val set stays frozen). Result: 1115 eps = 1037 train / 78 val.
norm_stats.json and the packed form of tasks/ live in armteam/hapticwam-teleop-dataset; the text_embeddings.pt prompt cache lives in armteam/hapticwam-teacher (never recomputed; the cache covers all 8 task keys incl. *_fail).

Episode format

Each episode directory holds one zarr group per stream (data + ts arrays, timestamps in master-clock seconds) plus meta.json. Streams:

Stream	Contents	Rate
`camera_scene_color`	JPEG images	~15 Hz
`tactile_left`, `tactile_right`	`wrench` (6D, N / N·m), `area` (mm²), `fields_ds` (72×96×8, f16), `keyframes` (144×192×8, f16), `infer_img` (288×384, u8)	~6.5 Hz
`arm`	`q`, `qd`, `tcp_pose`, `tcp_speed`, `ft`	~125–140 Hz
`gripper`	`pos`, `obj-detect`	~100 Hz
`actions`	Delta end-effector commands	~10 Hz
`actions_abs`	Absolute joint targets + gripper command	Event-based

Schema

One directory per episode: meta.json plus one zarr group per stream, each holding a data array (T, …) and a ts array (T,) of master-clock seconds. Every stream keeps its own T and its own timestamps — nothing is resampled onto a common grid — so align by nearest ts (observations) or first ts >= t (future targets).

Stream	Shape	dtype	Rate	Meaning
`tactile_{left,right}_fields_ds`	`(T, 72, 96, 8)`	f16	~5.7 Hz	tactile field stack, channels `[disp_x, disp_y, depth, shear_x, shear_y, fx, fy, fz]`; `depth` and the distributed force `fx,fy,fz` are in raw SDK units (uncalibrated)
`tactile_{left,right}_keyframes`	`(T, 144, 192, 8)`	f16	~2.5 Hz	the same 8 channels at higher spatial resolution, time-decimated
`tactile_{left,right}_infer_img`	`(T, 288, 384)`	u8	~5.7 Hz	the SDK's gel image (`getInferImg`)
`tactile_{left,right}_wrench`	`(T, 6)`	f32	~5.7 Hz	pad wrench `[Fx, Fy, Fz, Mx, My, Mz]`, N and N·m (SI: the SDK's 1e−2 N·m torques are scaled at the driver boundary)
`tactile_{left,right}_area`	`(T,)`	f32	~5.7 Hz	contact area, mm²
`arm_q`, `arm_qd`	`(T, 6)`	f64	125 Hz	joint position (rad), joint velocity (rad/s)
`arm_tcp_pose`, `arm_tcp_speed`	`(T, 6)`	f64	125 Hz	TCP pose `[x, y, z, rx, ry, rz]` (m + rotation vector, base frame) and twist
`arm_ft`	`(T, 6)`	f64	125 Hz	wrist wrench, N / N·m — on this CB3 arm a current-based estimate with a large pose-dependent bias, not a real F/T sensor
`gripper`	`(T, 2)`	f32	~95 Hz	`[position, obj]`: closure 0 (open) … 1 (closed), and the Robotiq gOBJ status 0..3 (`2` = stopped by contact while closing, i.e. holding)
`camera_scene_color`	`(T, 480, 640, 3)`	u8	15 Hz	scene RGB, JPEG-encoded per frame (quality 92) in a zarr `VLenBytes` object array; the group's attrs carry `encoding`, `frame_shape`, `frame_dtype`, `jpeg_quality`
`actions`	`(T, 7)`	f32	10 Hz	the canonical action: Δ-EE pose step `[Δx, Δy, Δz, Δrx, Δry, Δrz]` (m, rotation-vector rad) + commanded gripper closure
`actions_abs`	`(T, 7)`	f32	10 Hz	absolute command `[q_target(6) rad, gripper]`

Rates are what the rig actually achieved (the tactile SDK free-runs below its 8 Hz cap); use ts, never an assumed rate. A stream that produced no samples has no directory at all, so check before you read: a pad-free deploy (tag padfree:on) has no tactile_* streams, simulation exports and most deploy takes have no actions_abs (and sim adds contact_gt), and a re-derived policy rollout adds actions_plan (the executor's pre-clamp proposal).

meta.json carries task, text, operator, tags, policy, dagger_round, success, damage, notes, driver_modes, clock_calibration, config_hash, hardware_shapes, deploy_overrides, status and weight. It is authoritative for task and labels — directory names are not. status must be finalized for an episode to be trainable; success, the deliberate_failure tag and a _fail task name each zero the action-imitation loss.

Full schema — every field, unit, threshold, the time base and a runnable "read one episode" snippet — is docs/dataset_schema.md in the code repository.

Packaging

Loose — one directory per episode under tasks/<task>/, browsable in the file viewer. That is also why it is ~865,000 files: snapshot_download of the whole repo spends 1–3 hours enumerating before a byte of training data lands, which is why the training corpus is published packed as armteam/hapticwam-teleop-dataset (one .tar.zst per task, same episodes, plus samples/ for browsing).

hf download armteam/hapticwam-teleop-raw --repo-type dataset --local-dir . \
    --include "tasks/waffles/ep_waffles_1785592739_002/*"

# one episode, resolved to the single archive that holds it, from a code checkout
python tools/hub/fetch_episode.py --dataset teleop --episode ep_waffles_1785592739_002 --loose
python tools/hub/fetch_episode.py --dataset teleop --episode first --samples   # the sample

Tasks

task	episodes	role	median dur	median peak \|F\|	median contact
Carton	180	successes	17.4 s	12.9 N	17.9 mm²
Carton_fail	20	failure demos	17.7 s	26.3 N	36.1 mm²
waffles	180	successes	18.3 s	8.4 N	3.0 mm²
waffles_fail	20	failure demos	20.4 s	17.8 N	25.5 mm²
egg	180	successes	27.6 s	8.8 N	18.0 mm²
egg_fail	20	failure demos	18.4 s	33.5 N	49.1 mm²
whiteboard	180	successes
whiteboard_fail	10	failure demos

(tasks/ view = 790 eps: 712 train / 78 val; val = last 2 sessions per success task.)

batch_20260822 (`archive/20260822_*`, not yet in `tasks/`)

task	eps	labels
Carton, egg, waffles, whiteboard	70 each	`success=true`, tags `[full, batch_20260822]` — ordinary demos, appended as train
Carton_fail, egg_fail, waffles_fail	15 each	`success=false`, `failure_demo=true`, tags `[full, deliberate_failure, undergrasp, batch_20260822]` — under-grasp: gripper closed on little/nothing, then the task was continued as if holding ("phantom carry"); actions are never imitated (`action_weight=0`), tactile/contact/event heads still train on them

Notes

success=true everywhere in the success tasks. *_fail tasks are DELIBERATE failure demonstrations: v4 *_fail (over-squeeze / induced slip, higher forces/contact) carry success=true ("the episode captured the intended failure"); batch_20260822 *_fail (under-grasp) carry success=false. Training treats both identically — is_failure_demo() (phantom/data/schema.py) fires on success=false OR the deliberate_failure tag OR a task ending in _fail, and zeroes the action-imitation loss for that episode.
Tags: full = complete teleop take (recorder default); batch_<YYYYMMDD> = intake batch (provenance only, no training effect); deliberate_failure / undergrasp = failure-demo kind.
tactile_contact=false marks episodes where the grasp landed outside the sensor pads (valid vision/proprio demos, no tactile signal).
Split rule: per success task, the last 2 sessions (chronological) are val, the rest train; failure demos are train-only. Re-split freely via the manifests — they are the source of truth, not the folder layout.
Peak forces briefly exceed the 30 N pad ceiling in a handful of episodes (dynamic spikes, mostly failure demos) — flagged in quality_notes.

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` ← you are here	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`	dataset	per-take evidence behind the paper's tables — scored CSVs, probe JSONs, figures