armteam/hapticwam-teleop-dataset

HapticWAM teleoperation dataset

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.

1,115 teleoperated manipulation episodes with fingertip tactile sensing, packed as per-task .tar.zst shards. This is the teleoperation corpus the HapticWAM teacher is trained on.

Hardware: UR3 + Robotiq 2F-85 + 2x Daimon DM-Tac W2L fingertip sensors + RealSense scene camera, teleoperated through an Echo exoskeleton leader.

Related repos: policy rollouts live in armteam/hapticwam-rollouts; the raw as-recorded session tree is armteam/hapticwam-teleop-raw; models are under armteam/hapticwam-teacher, -student, -baselines, -ablations.

Contents

File	Size	Episodes
Carton.tar.zst	14.53 GB	180
Carton_fail.tar.zst	1.50 GB	20
egg.tar.zst	21.59 GB	180
egg_fail.tar.zst	1.70 GB	20
waffles.tar.zst	14.92 GB	180
waffles_fail.tar.zst	1.80 GB	20
whiteboard.tar.zst	15.33 GB	180
whiteboard_fail.tar.zst	0.65 GB	10
batch_20260822.tar.zst	30.65 GB	325
manifests.tar	1.2 MB	index over the 790 task episodes
manifests_v6.tar	1.2 MB	index over all 1,115 episodes
norm_stats.json	42 KB	action/state normalisation statistics
total	102.67 GB	1,115

MANIFEST_OF_RECORD.md says which manifest to use. Read it before training.

Unpacking

The eight task shards unpack directly into the canonical flattened view:

hf download armteam/hapticwam-teleop-dataset --repo-type dataset --local-dir .
mkdir -p data/episodes/tasks
for f in Carton egg waffles whiteboard Carton_fail egg_fail waffles_fail whiteboard_fail; do
  tar -I zstd -xf "$f.tar.zst" -C data/episodes/tasks
done
tar -xf manifests_v6.tar -C data/episodes      # -> manifests/all.jsonl (1,115 rows)

batch_20260822.tar.zst unpacks to the raw session layout (20260822_<time>_<label>/ep_*), 34 sessions / 325 episodes, not the tasks/<task>/ layout. It is normalised and appended to the task view at provision time; the v6 manifest already carries the resulting rows and paths.

The shards were made as one archive per task because the loose per-episode layout runs to roughly 865,000 files, well past the hub's file-count guidance, and enumerating it before training costs hours. A shard is a plain tar compressed with zstd; nothing else is needed to read it.

Episode format

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

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

Manifest rows carry success, split (train/val), tactile_contact, duration_s, peak_force_N, max_contact_mm2, failure_demo, session, operator, task, text and path. manifests/quality_full.csv holds the full per-episode quality audit with per-stream sample counts, rates and gaps.

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

• Packed. One <task>.tar.zst per task (plain tar + zstd), members <task>/ep_<task>_<epoch>_<idx>/…, so a shard unpacks straight into the canonical tasks/ view. batch_20260822.tar.zst unpacks to the raw session layout 20260822_<time>_<label>/ep_*/… instead.
• The episode → shard map is the task, which the episode id carries and manifests_v6.tar (manifests/all.jsonl, one row per episode with path, task, success, split, …) confirms.
• index.jsonl is a per-file index of this repository, one JSON object per file with the columns path, size, sha256 (the LFS hash; null for small non-LFS files), blob_id and source (where the file was copied from in the 2026-09 restructure). Use it to verify a download — it does not map episodes to shards.
• samples/<task>/ep_*/ holds four complete episodes, one per task, extracted verbatim from the shards (see samples/README.md) — browse the format without a 15 GB download.
• The same episodes exist loose, one directory each, in armteam/hapticwam-teleop-raw under tasks/<task>/.

tar -I zstd -xf waffles.tar.zst -C tasks waffles/ep_waffles_1785592739_002   # one episode

# 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
python tools/hub/fetch_episode.py --dataset teleop --episode first --samples   # the sample

Tasks

task	episodes (v4 view)	+batch_20260822	role	median dur	median peak |F|	median contact
Carton	180	+70	successes	17.4 s	12.9 N	17.9 mm²
Carton_fail	20	+15	failure demos	17.7 s	26.3 N	36.1 mm²
waffles	180	+70	successes	18.3 s	8.4 N	3.0 mm²
waffles_fail	20	+15	failure demos	20.4 s	17.8 N	25.5 mm²
egg	180	+70	successes	27.6 s	8.8 N	18.0 mm²
egg_fail	20	+15	failure demos	18.4 s	33.5 N	49.1 mm²
whiteboard	180	+70	successes
whiteboard_fail	10	0	failure demos
total	790	+325				1,115

The 790-episode view splits 712 train / 78 val. With batch_20260822 appended the corpus is 1,115 episodes = 1,037 train / 78 val; the val set stays frozen across both views.

batch_20260822

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]

The batch failure demos are under-grasps: the gripper closed on little or nothing and the task was then continued as if the object were held ("phantom carry"). Their actions are never imitated (action_weight=0); the tactile, contact and event heads still train on them.

Notes

• success=true throughout the success tasks. The *_fail tasks are deliberate failure demonstrations. The v4 *_fail episodes (over-squeeze / induced slip, higher forces and contact areas) carry success=true, meaning "the episode captured the intended failure"; the batch_20260822 under-grasp episodes carry success=false. Training treats both the same way: a failure demo is anything with success=false, a deliberate_failure tag, or a task name ending in _fail, and its action-imitation loss is zeroed.
• 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. These are valid vision and proprioception demos with no tactile signal.
• Split rule: per success task, the last two sessions chronologically are val and the rest train; failure demos are train-only. Re-split freely from 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). These are flagged in quality_notes.
• One recorder configuration (config_hash b2504b6d5ff0c29d) covers the whole corpus, and every episode here is policy: teleop, dagger_round: -1 — pure human teleoperation.
• index.jsonl lists every file in this repo with its size and LFS sha256, and the path it was copied from.

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 ← you are here	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	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

Data: CC-BY-4.0. The HapticWAM code and the model weights in the repos above: Apache-2.0.