NVIDIA releases post-training framework for real-time interactive Cosmos world models
chenhsuanlin · x · 2026-10-22
NVIDIA released a post-training framework for real-time interactive world simulation for Physical AI, built on Cosmos 3 world foundation models. Key points:
- Achieves real-time interactivity via causal training, distillation, and optimized streaming inference
- Post-trained models stream visual observations through autoregressive generation, responding to incoming actions and scene controls in real time
- Ships example checkpoints for robotics, autonomous vehicles, and sim-to-real transfer
- Includes optimized streaming inference code
The authors note that while Cosmos provides broad visual knowledge, each simulation setting has its own controls, data, and evaluation needs — the shared recipe and example models give researchers a practical starting point for building specialized interactive world simulators.
Related event: NVIDIA Unveils Cosmos Real-Time Interactive World Simulation(4 posts)→
More from Embodied
- 158 of 188 RFD4-Proteina-designed TrkA binders hit sub-micromolar affinity in the lab — AllThingsApx · 2026-10-22
- Team spends $5K benchmarking GPT, Claude, Gemini on LIBERO-10 — best success rate just 44% — YuXiang_IRVL · 2026-10-22
- NVIDIA's Boris Ivanovic talks embodied reasoning models for physical AI at Mila seminar — GlenBerseth · 2026-10-22
- Upstream CEO: hacked robots can turn dangerous without losing control — TechNadu · 2026-10-22
- Xpeng's He Xiaopeng: IRON humanoid robot's natural interaction starts with simple moments — Scobleizer · 2026-10-22
- Stanford opens its robot autonomy course materials: free PoRA textbook and lectures for everyone — drmapavone · 2026-10-22