VLA vs. world models: a hands-on tutorial builds four robot paradigms on the same data
GlenBerseth · x · 2026-10-19
Glen Berseth's team is running a hands-on tutorial at MAINeducational in Montréal (Oct 26) that builds and compares four robot-learning paradigms on identical robot data:
- Policy (VLA): predicts actions only, trained with imitation loss — never sees what happens next
- Pixel world model (Dreamer-style): predicts the next frame, reconstruction loss in pixel space
- Latent world model (V-JEPA-style): skips pixels, predicts the next embedding
- World action model (Cosmos Policy): predicts both actions and the future, letting future prediction shape behavior
A shared chart breaks down each model's inputs, outputs, and loss spaces around the core question: what should a robot model actually learn to predict?
Related event: Montreal Workshop to Compare Four Robot Modeling Paradigms on Same Data(3 posts)→
More from Embodied
- Smart glasses test: Octen search at 60ms, model generation is the bottleneck; Mercury 3 wins — Scobleizer · 2026-10-20
- Cheap Robot Arms Are Turning 3D Printers Into Overnight Factories — Scobleizer · 2026-10-20
- Dream-1 claimed to score 1.3x higher than GPT-6 Astra Ultrafast at 1.7x speed — gharik · 2026-10-20
- Ex-NVIDIA team launches Dream Labs with Dream-1 robot model via public API at $3/hour — gharik · 2026-10-20
- Dream-1 claims new Pareto frontier for robot control: 16x cheaper, 1.7x faster, 59 vs 48 on RoboDojo — scott_e_reed · 2026-10-20
- Surgeons Complete 100+ Cataract Surgeries Using Apple Vision Pro and ClearSphere App — asymco · 2026-10-20