Action-conditioned causal rollouts
Five observation streams. Recorded actions. Eight observed frames, followed by short predictions or free-running trajectories.
How to read these results
Each video shows GT above prediction. Columns are main view, left wrist, right wrist, left tactile, and right tactile. The first eight frames are observed history; the rest are predictions. GT visuals are decoded cached latents.
Contact means max(left force, right force) > 0.5 N. None: no contact in the initial 16-frame window. Full: contact in at least 80% of that window. Onset/release: a contact transition in the eight predicted frames. Long rollouts may change contact state after the initial window.
Scores cover future frames only. Long scores average the 15 generated windows; the table averages the currently displayed cases equally. PSNR and SSIM are higher-is-better; LPIPS is lower-is-better. Cases are selected for contact coverage, not as a representative full-test benchmark. Dataset scores are not a controlled comparison.
Generation uses 35 UniPC steps per causal latent block, seed 0. Long rollouts reuse generated observations as history and retain recorded action conditions. Motherboard 0909 uses an older force calibration than 0911/0912. PushT episode_006 in the requested split is called episode_007 in the current data.
Download selections and metrics (JSON)Per checkpoint — Motherboard (including custom attn epochs 002/004/010): 20 short + 20 long · Rope: 34 short + 34 long · PushT epochs 003/007: 15 short + 14 long; epoch 013: 30 short + 28 long across its full eight-segment test split. Motherboard and Rope reuse their earlier selected windows. PushT epoch 013 includes the earlier windows plus additional test segments.