Tactile world model · contact supervision

Fixing the "predict no-contact" bias

p01rand (dense MSE only) under-predicts contact — val area 0.29x GT. contactw1 adds a strong contact-region reconstruction loss (masked MSE, w=1.0, threshold 0.20, no false-contact term) then recovers contact. Both are full 3-view + 2-tactile models.

runcontact IoU (val) pred/GT area (val)pred intensity in contact
p01rand0.1730.295.84
contactw10.2020.857.04
Contact maps (matched windows)

Columns: GT tactile RGB · GT contact · p01rand pred · contactw1 pred (red = above tau=8 contact threshold).

Short-horizon rollouts (2.7s) — p01rand | contactw1

Matched val windows. Click the play button or scrub; the view clip is the clock, tactile clips follow.

window 0

2026-05-10__episode_005.pt · t=604 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 1

2026-05-10__episode_005.pt · t=606 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 2

2026-05-10__episode_006.pt · t=1116 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 3

2026-05-10__episode_006.pt · t=1226 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 4

2026-05-10__episode_006.pt · t=1228 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 5

2026-05-10__episode_006.pt · t=1238 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 6

2026-05-10__episode_006.pt · t=1420 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 7

2026-05-10__episode_006.pt · t=2460 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 8

2026-05-10__episode_006.pt · t=2496 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

window 9

2026-05-10__episode_006.pt · t=2502 — GT ‖ prediction. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
Long-horizon rollouts (40s AR) — p01rand | contactw1

Full 40-second autoregressive rollout on held-out episodes — tests whether the recovered contact prediction stays stable over long horizons (it does; no collapse/drift).

40s · 2026-05-10__episode_005.pt

Full 40-second autoregressive rollout. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16

40s · 2026-05-10__episode_006.pt

Full 40-second autoregressive rollout. top = middle view, then tactile left / right.

p01rand
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16
contactw1
middle view GT · pred
tactile · left GT · pred
tactile · right GT · pred
0/16