Turing Post weekly: RoboJEPA, Long-WAM, VioLA and why EmbeddingGemma 2 matters
TheTuringPost · x · 2026-10-14
Turing Post's free weekly digest (FOD#171) covers two main threads:
Models to know this week
- RoboJEPA: latent world models trained across 12 robotic embodiments, with prediction accuracy linked to planning performance.
- Long-WAM: a world-action model using longer visual histories for better robot control.
- VioLA: a generalist humanoid control model learning body and hand movements mostly from human recordings.
- UNITAS: a 3D-native world-action model (truncated in source).
Why EmbeddingGemma 2 is interesting
Google's newly released EmbeddingGemma 2 is a 740M-parameter multimodal embedding model that runs on-device under Apache 2.0. It can find videos from spoken descriptions, search audio with written questions, and retrieve images from sketches, unifying text, images, audio and video in one embedding space. The author argues tiny embeddings are key to building agent memory. The issue also previews O'Reilly's free Agent Memory online conference and upcoming pieces on 3D Gaussian Splatting and Looped Transformers.
Related event: Turing Post Weekly: Robotics World Models Take Center Stage(3 posts)→
More from Embodied
- DynaRobotics' robot Taku rings the Nasdaq opening bell — chris_j_paxton · 2026-10-14
- VAP-TAMP wins IROS best paper: VLMs help robots handle surprises — chris_j_paxton · 2026-10-14
- NAVER's WATCH: Cross-Video Human-Video Pretraining Lifts Robot Success Rate to 77.7% — naver-ai · 2026-10-14
- LeRobot adds remote inference via Zenoh: robot stays light, policy runs on cloud GPU — kamathsblog · 2026-10-14
- Robot Dog Tries to Climb Its Own Back: Conitzer's Latest Funny AI Fail — conitzer · 2026-10-14
- This Week in New Models: RoboJEPA, Long-WAM, VioLA and 4 More to Know — TheTuringPost · 2026-10-14