HapticWAM baselines
The three baseline policies the HapticWAM student is compared against, at the exact steps the paper deploys, plus the evaluation sweeps that chose those steps.
| Folder | Files | Size | Baseline | Deployed checkpoint |
|---|---|---|---|---|
diffusion_100k/ |
7 | 1.05 GB | Diffusion Policy | pretrained_model/ at 100k steps |
xvla_20k/ |
7 | 1.76 GB | X-VLA 0.9B | pretrained_model/ at 20k steps |
pi05_phantom_expert_v1/ |
18 | 9.69 GB | pi0.5 expert fine-tune | 020000/pretrained_model/ |
eval_baselines/ |
10 | 1.2 MB | evaluation JSONs, see below | |
eval_pi05/ |
11 | 1.2 MB | pi0.5 step-selection sweep |
index.jsonl lists every file here with its size, LFS sha256 and the path it was copied from.
What the eval files cover
eval_baselines/ holds the diffusion-policy sweep (eval_dp_025000 … eval_dp_100000) and the
X-VLA sweep (eval_xvla_005000 … eval_xvla_020000), which is how 100k and 20k were chosen.
It also holds A_visiononly_nfe1.json and B_nodistill_nfe1.json. Those two score the
vision-only and no-distillation Cosmos arms, whose checkpoints live in
armteam/hapticwam-ablations rather than here; they are kept together because they are one
baseline table in the paper.
eval_pi05/ holds the step sweep for the pi0.5 expert. The deployed pi0.5 arm is step
20000 (pi05_020000_terminal.json). The eval_resume_* files score a run that was resumed
to 60k steps; that checkpoint is not deployed and lives in armteam/hapticwam-ablations as
pi05_phantom_expert_v1_resume60k/.
pi05_phantom_expert_v1/ also carries 020000/training_state/ (optimizer and scheduler state,
2.2 GB) and paligemma_tokenizer/, both needed to resume or to run the policy.
Data
All three baselines were trained on
armteam/hapticwam-teleop-dataset.
Models: armteam/hapticwam-teacher,
armteam/hapticwam-student.
Licence
Apache-2.0.