本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://news.mit.edu/2024/precision-home-robotics-real-sim-real-0731
Summary
MIT CSAIL’s RialTo system inverts the traditional sim-to-real pipeline: instead of building a simulator first, it scans real home environments with a phone (using NeRFStudio) to create high-fidelity digital twins, trains policies via GPU-parallel RL in those twins, then deploys to real robots. The result is 67% higher task success compared to traditional imitation learning, with training time compressed from weeks to three hours.
MIT CSAIL 的 RialTo 反轉傳統 Sim-to-Real 流程：先用手機掃描真實家庭環境建立數位孿生，再於孿生中平行訓練，最後遷移到真實機器人。相比傳統模仿學習提升 67% 成功率，並在廚房、書架、餐具等家庭任務中驗證了在視覺干擾和物理擾動下的穩健性。
Key Points
- “Real-to-sim-to-real” inverts the traditional pipeline: scan first, simulate second, deploy third
- Phone scanning + NeRFStudio generates deployable 3D scenes for Isaac Sim in minutes
- GPU-parallel RL collects equivalent of “100 days” of experience in ~3 hours
- 67% improvement in task success over imitation learning on household manipulation tasks (opening toaster, placing books, kitchen utensils, drawers/cabinets)
- Robust to object position randomization, visual distractors, and physical perturbations
- Limitation: still needs initial human demonstrations; deformable objects and liquids remain hard; training takes ~3 days total
Insights
The key insight is accessibility: the barrier to creating a simulation environment is not specialized tools but the difficulty of scanning and reconstructing the exact deployment environment. RialTo makes this a 10-minute phone task instead of a weeks-long CAD modeling effort, which fundamentally changes who can use sim-to-real techniques.
The “personalization” angle is significant: a robot trained in your specific kitchen, with your specific drawer positions and cabinet heights, will dramatically outperform one trained on generic kitchen environments. This points toward a future where per-household fine-tuning is standard.
Connections
Raw Excerpt
“Anyone with a phone can scan their own home environment, train a robot policy optimized for that specific environment, and deploy it — democratizing embodied AI beyond the laboratory.”