What should a robot model predict? Four paradigms compared on the same robot data
GlenBerseth · x · 2026-10-19
Montreal researcher Glen Berseth poses a hands-on tutorial question: what should a robot model actually predict? He will build and compare four paradigms on the same robot data:
- VLA policy: predicts actions directly, never seeing what comes next
- Pixel world model (Dreamer-style): imagines the next frame
- Latent world model (V-JEPA): predicts the next embedding, skipping pixels
- World action model (Cosmos Policy): predicts both actions and the future
The tutorial runs Oct 26 at the MAIN educational event in Montreal — a rare chance to see the four leading robot modeling paradigms compared side by side on identical data.
Related event: Montreal Workshop to Compare Four Robot Modeling Paradigms on Same Data(3 posts)→
More from Embodied
- AI Lab Offers Free Plumbing and HVAC Repairs to Film Wearable-Sensor Video for Robot Training — whatdowecallu · 2026-10-20
- ETH Zurich teaches a robotic hand to walk on its fingers with reinforcement learning — MIT_CSAIL · 2026-10-20
- LeCun co-authored LeHami: 1830s symplectic physics fixes world model collapse, 2.8x more accurate with 1/8 compute — alex_verem · 2026-10-19
- MolmoAct2 lead Duan Jiafei joins NUS with multimillion-dollar robotics grant — DJiafei · 2026-10-19
- PixVerse R2 acts like a world model: WASD movement and text inputs reshape a live game world — hey_abusiddik · 2026-10-19
- Robotics writer admits he was wrong: Physical Intelligence just shipped a model API — binarybits · 2026-10-19