T2-z Optimizes Environment Latents at Test Time for Robust Robotic World Models
yining_hong · x · 2026-10-17
T2-z conditions robotic world models on optimizable environment latents to capture hidden factors like friction and mass, trained purely via world prediction without supervision on factor identity, values, or count. Notably, PCA on learned latents reveals a 2D structure matching the underlying mass×friction distribution. Validated across 7 simulated and 4 real-world robot settings.
More from Embodied
- Startup Elorian launches Glimpse, a visual thinking model that could automate robot data labeling — LinusEkenstam · 2026-10-17
- Sandbar's Stream Ring nails whisper-quiet dictation across iOS and Mac in demo — thesephist · 2026-10-17
- RAX open-sourced: frozen coding agents learn robot tasks via online exploration, zero demos — du_yilun · 2026-10-17
- U. Tokyo's TwinX robot: one wire links two wheeled-legged bots for morphing and teamwork — LexiLove · 2026-10-17
- Snapchat Specs pop-up in LA: tech is cool but the $2k glasses 'look ridiculous' — SimonBalmain · 2026-10-17
- IEEE profiles roboticist Peter Stone: robot soccer, service robots and being a generalist — PeterStone_TX · 2026-10-17