armteam/hapticwam-ablations

HapticWAM ablations and supporting runs

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.

Every training arm that is not the deployed teacher, the deployed student or a headline baseline, plus the complete evaluation sweeps for all of them. Nothing here is deployed on the rig. This repo exists so that each number in the offline tables has a checkpoint behind it.

Deployed models live in armteam/hapticwam-teacher and armteam/hapticwam-student; headline baselines in armteam/hapticwam-baselines.

Teacher arms

Folder	Files	Size	What it is
teacher_v5_ftA/	7	2.36 GB	v5 fine-tune A, steps 500–3000, log train_ftA.log. The teacher the ftA student line was distilled from; it backs the earlier offline table.
teacher_v6_ftA/	2	0.39 GB	v6 fine-tune A, step 1500, plus GO_STUDENT.json.
teacher_v6_ft_video1p0/	5	1.57 GB	v6 with the video loss weighted 1.0, steps 500–2000. The world-model loss ablation.
teacher_v6_ft_control/	4	1.57 GB	v6 fine-tune control arm for the world-model loss study, steps 500–2000. The reference the other two are read against.
teacher_v6_ft_noVideoLoss/	4	1.57 GB	v6 with the video loss switched off, steps 500–2000. Isolates the contribution of the video loss term.
teacher_v6_ft_videoAttend/	4	1.57 GB	v6 with video attention enabled, steps 500–2000. The attention variant of the same study.
teacher_v6_simft_multitask/	5	1.57 GB	multitask sim fine-tune, steps 1000/1500/2000/3000, log train.log.
teacher_v5_batch0822/	1	0.39 GB	teacher_003000.pt, the v5.6 checkpoint the FT-A fine-tune starts from. The provisioning scripts download it as the FT-A initialisation. Only this one file was copied forward; the rest of that v5 probe sweep was not.

Student and control arms

Folder	Files	Size	What it is
hid_mt/	6	1.49 GB	student distilled from the multitask teacher, steps 250–1000, logs distill_B.log, extend_B.log.
hid_2k_ftA/	5	1.49 GB	ftA student, round 2, steps 500–2000, log distill_r2.log.
hid_r2_ftA/	5	1.49 GB	ftA student, round 2. Same lineage, same sizes and the same log name as hid_2k_ftA; the two are kept apart because the hub cannot tell which the earlier table quoted. Treat them as one arm until that is resolved.
hid_v6_r2/	5	1.49 GB	v6 student, round 2, steps 500–2000, log stage3_v6.log.
ctrl_v6/	2	0.37 GB	the v6 no-distillation control, teacher_001200.pt, log distill_v6.log.
ctl_ftA_2k/	1	0.37 GB	the step-matched no-distillation control, teacher_002000.pt: the same 2,000 steps as the deployed teacher, trained without distillation, so the comparison isolates distillation from training length. It backs the step-matched control row of the offline table.
cosmos_nodistill_v1/	1	0.37 GB	Cosmos arm without distillation, teacher_006000.pt. Scored by eval_baselines/B_nodistill_nfe1.json.
cosmos_visiononly_v1/	1	0.37 GB	Cosmos arm, vision only, teacher_006000.pt. Scored by eval_baselines/A_visiononly_nfe1.json.
pi05_phantom_expert_v1_resume60k/	7	7.47 GB	the pi0.5 expert resumed to 60k steps. Not the deployed pi0.5 arm — that is step 20000, in hapticwam-baselines. Kept because eval_pi05/eval_resume_*.json scores it.

Evaluation sweeps (complete)

Folder	Files	Covers
eval_r2/	40	every round-2 student rung: 2k_student_*, r2_student_*, v6r2_student_*, nowrist_student_*, nowrist_cont_student_*. JSON plus log each.
eval_mt/	16	hid_mt_student_* and hid_simft_student_* at steps 250–1000.
eval_v6/	14	v6_student_student_* at steps 500–2000, v6_control, and the v6 teacher at NFE 1 and 5.
eval_abl/	8	teacher_v6_ft_video1p0 at steps 500–2000, NFE 1.
eval_0906/	6	the 2026-09-06 ftA and v5.6 comparisons.
eval4/	2	ftA and v6_ftA teachers at NFE 1.

These folders are the canonical, complete sweeps. hapticwam-teacher, -student and -baselines each carry a small copy of the subset that scores their own checkpoints; where the two disagree in future, this repo is the one to trust.

index.jsonl lists every file here with its size, LFS sha256 and the path it was copied from.

Not here

Superseded rounds — distillation rounds 0 and 1, the v2/v3/v4 teachers, the round-0 controls and the v5 probe sweep — were not copied forward and are not published.

ctl_ftA_2k/, teacher_v6_ft_control/, teacher_v6_ft_noVideoLoss/ and teacher_v6_ft_videoAttend/ are the folders here that were not copied from another hub repo: they had never been uploaded anywhere and existed on a single lab machine. Their index.jsonl rows therefore record a filesystem path rather than a hub path as their source. The three teacher_v6_ft_* arms above were uploaded on 2026-09-20, each verified by size and LFS sha256 against the machine copy before that copy was cleared.

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 ← you are here	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	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

Model weights and the HapticWAM code: Apache-2.0. The datasets listed above: CC-BY-4.0.