Amazon's DreamFAR: Real-Time Robot World Model at 10+ FPS, Playable Online
CyberRobooo · x · 2026-10-30
Amazon FAR, with Cornell, Maryland, and UPenn, released DreamFAR, a real-time interactive world model for robotic manipulation, with a live demo (mouse, keyboard, webcam hand, or leader arms).
Key numbers: 10+ FPS real-time interaction, 5+ minute autoregressive rollouts, 1000+ tasks with diverse objects—articulation, deformables, tool use, cloth, particles, paper folding, cables, writing, and more.
Method:
- Latent flow-matching world model: a tokenizer encodes each camera view into compact latents; a causal transformer predicts the next latent from actions and recent latents
- Contact-based memory: keyframes inserted on contact make/break, indexed by end-effector pose—solving object-state consistency over long interactions
- Streaming transformer design; its own predictions become its history, so one network simulates recorded actions, policies, or human operators
Code coming soon.
More from Embodied
- Robotics Startup Machenta Debuts Instant Robot Learning, Hiring in London — Ed__Johns · 2026-10-30
- Home-cleaning robot charging $30/hour points to a multi-billion-dollar TAM — xmercury_one · 2026-10-30
- Genie 4 sparks debate: robots' GPT moment nears, home humanoid adoption doesn't — FuSheng_0306 · 2026-10-30
- Handwriting Robot Completes Biology Homework With Cursive Text and Ink Diagrams — alvelda · 2026-10-30
- Meta's RoboJEPA: First Scaling Laws for Multi-Embodiment Robotic World Models — apsarathchandar · 2026-10-30
- Foldable iPhone Duo already in stock at Paris Apple Store: 512GB at €2,589, pickup today — mitchdeg · 2026-10-30