Tactile Diffusion-Forcing world model · tac_s5_hf (final)
Where the tactile world model works, and where it fails
The same model evaluated across eight conditions — train vs held-out,
short (1.3 s) vs long (16 s) horizon, and high vs low contact — with
three examples each. Every clip is ground truth on the left, model prediction on the right. The point of
the grid is that a single headline number hides three very different regimes.
The short answer. Short-horizon prediction genuinely works and generalizes. Long-horizon
rollout on held-out data is worse than doing nothing — a baseline that simply freezes the last
context frame beats it. And in every high-contact condition the model erases fine texture that was
already present in its own context.
The comparison baseline
Every cell is scored against frozen context: take the last ground-truth context frame and
hold it for the whole prediction horizon. It is the "predict no change" null model. A world model that
cannot beat it is not contributing anything on that condition.
PSNR and texture are measured on predicted frames only (decoded frames 5 onward; frames 0–4
come from the ground-truth context latents). Texture = mean Sobel gradient energy of the prediction divided
by that of the ground truth, so 1.0 means the prediction is as detailed as reality and lower means blur.
All eight conditions
condition
contact
model PSNR
frozen PSNR
Δ
model texture
frozen texture
beats frozen
Read the contact column before comparing splits. The held-out episodes never reach the
contact levels the training set does — "high contact" means ~0.22–0.38 on test but up to 0.68 on train.
Only one episode in the whole 36-episode corpus (0511_episode_017, contact 0.435
vs a 0.160 median) is genuinely high-contact, and it is in the training split. Train-vs-test rows in the
high-contact conditions are therefore not matched comparisons.
What the three regimes mean
Short horizon — real, generalizing skill. The model beats frozen context on essentially
every short window, on held-out data as much as on training data. Given 8 frames of context it predicts
the next 8 frames of gel deformation better than assuming nothing moves. This is the part that works.
Long horizon on held-out data — net harmful. Frozen context wins on every held-out
16 s segment. Notably the deficit is not accumulating drift: the model loses about 10 dB at the
very first predicted frame and is then flat, decaying only 1 dB across 16 s while frozen context decays
7.7 dB as the scene genuinely changes. A constant offset that does not compound points at a systematic
first-step mismatch rather than error accumulation.
High contact — texture is destroyed, not merely missed. On contacting sensors the model's
texture energy falls below frozen context. Freezing a stale frame preserves more fine structure
than the model predicts. Ridges and pin-dot patterns that are clearly visible in the model's own context
frames are gone within about five predicted frames. Separately, contact onset — a new object
entering the gel — is never predicted, which is likely not learnable from proprioception alone and is the
case for conditioning on vision.
Caveats
Three examples per cell. This grid is diagnostic and illustrative, not a powered statistical test.
The validation split is two episodes, 908 of 47,483 windows (1.9%). Everything labelled "test" rests
on that.
The 16 s held-out numbers carry an unresolved caveat: latents are independent per-window VAE encodes,
so the long-rollout ground truth is stitched from different encodes than the one the model continues
from. That may account for part of the first-frame offset. Unverified — treat long-horizon held-out
magnitudes as provisional. The short-horizon results do not depend on it.
Contact fraction comes from reference-subtraction masks; it measures pressure area, which is not the
same as visible imprint texture. Some windows with moderate contact carry little visible structure.