NVIDIA shows one post-training recipe turns Cosmos 3 into real-time world simulators for robots and AVs
liu_mingyu · x · 2026-10-22
NVIDIA released new Cosmos updates demonstrating how a single post-training recipe turns Cosmos 3 into real-time, streaming interactive world models for robotics and autonomous driving.
Key points:
- Cosmos3-Nano-Sim-Bimanual predicts visual dynamics from camera navigation and robot actions, streaming next observations as new controls arrive; it beats DreamDojo 2B/14B on video quality and reference-frame PSNR across AgiBot-World and AgiBot-GEAR.
- Cosmos3-Nano-Sim-Transfer converts depth, segmentation, edges, or blurred video into detailed RGB observations, with language-controlled appearance while the simulator maintains physical state.
- Technical report, model cards, and GitHub code are available.
More from Embodied
- Tesla Hits Record 75% FSD Attach Rate in North America, Up From 55% in Q2 — bennash · 2026-10-22
- OpenTac team to demo tactile sensors at CoRL with small-scale sensor giveaway — anand_bhattad · 2026-10-22
- Kodiak Robotics to start humanless trucking 'within weeks,' already hauling for IKEA with safety drivers — Scobleizer · 2026-10-22
- RAX Uses a Frozen Coding Agent to Drive Robots: 96.7% on DexJoCo at 1/20 the Cost — du_yilun · 2026-10-22
- CMU to Launch Hands-On Humanoid Learning Course in Spring 2027, Fully Open-Sourced — GuanyaShi · 2026-10-22
- S1 data glove captures demos of Rubik's Cube solves, card shuffling, butterfly knife tricks — Haoyu_Xiong_ · 2026-10-22