New paper from Robbyant, with inference code and checkpoints already public.

LingBot-Touch is a self-supervised tactile encoder trained JEPA-style: it predicts latent representations of masked contact instead of reconstructing raw sensor signals. The paper tests it against same-size MAE and VAE reconstruction models on the same data. Tactile-to-video top-5 recall at 1K hours: 32.1% (JEPA) vs 5.6% (MAE) vs 4.4% (VAE). At 4K hours JEPA reaches 49.3% and MAE stays under 9%, while the VAE actually gets worse, down to 3.1%. Their explanation: most of a tactile clip is near-zero background, so reconstruction loss can shrink without learning anything about actual contact.

Two design choices stood out. The masking ratio warms up from 40% to 80% over 300 steps, because starting at high masking hid every contacting taxel in sparse signals and caused gradient explosions.…