# NVIDIA Demonstrates Sim-Trained Robots in Real World > Source: > Published: 2026-05-28 19:39:13.850323+00:00 # NVIDIA Demonstrates Sim-Trained Robots in Real World NVIDIA Research presented eight new papers at the International Conference on Robotics and Automation (ICRA) showing progress on simulation-to-real transfer for robotics, the NVIDIA blog reports. The work covers navigation, multi-arm coordination, grasping, assembly, and vision-language-action reasoning. Interesting Engineering and NVIDIA describe specific systems: the COMPASS framework, trained inside **NVIDIA Isaac Lab** simulations, reportedly achieved about **80 percent** success across **20** real-world navigation trials (Interesting Engineering). The team also reports a real-time grasp controller, **Grasp-MPC**, and a GPU-powered multi-arm planner, **ScheduleStream**, which the NVIDIA blog reports produced a **3x** speedup in multi-arm planning scenarios and has code available. Together, the papers aim to narrow the sim-to-real gap and validate simulation-trained policies on physical robots. ### What happened NVIDIA Research presented eight new papers at the International Conference on Robotics and Automation (ICRA) that demonstrate simulation-trained policies transferring to physical robots, the NVIDIA blog reports. The blog states the set of papers tackles perception, planning, and control across dynamic environments. Interesting Engineering reports that the COMPASS framework, trained in **NVIDIA Isaac Lab** simulations, achieved about **80 percent** success across **20** real-world navigation trials. Interesting Engineering also describes a real-time grasp controller called **Grasp-MPC** that refines grasps as the robot approaches objects. The NVIDIA blog reports **ScheduleStream**, a GPU-based multi-arm planning framework, yielded a **3x** speedup in parallel arm planning and that code for the framework is available. ### Technical details The sources describe work spanning multiple technical areas: - • **multi-arm coordination**(ScheduleStream), - • **body-agnostic navigation and generalization**(COMPASS), - • **grasping in clutter**(Grasp-MPC), - • **vision-language-action reasoning** for task-level planning, per the NVIDIA blog. The blog emphasizes GPU-accelerated simulation and planning as a common infrastructure thread; the reporting links measured outcomes (success rates, runtime speedups) to these systems' ability to run larger, faster simulated training and planning workloads. ### Industry context Editorial analysis: Companies and research groups working on sim-to-real routinely combine high-fidelity simulation, domain randomization, and GPU-accelerated training to scale policy training. Observers note that measurable real-world validation, such as **80 percent** success over multiple trials or explicit runtime speedups, is the primary metric practitioners use to judge whether a sim-to-real approach is practically useful for deployment scenarios. ### What to watch Editorial analysis: Observers should track code releases, benchmark protocols, and reproducibility across robot bodies and labs. Key signals will be independent replications of COMPASS-style body-agnostic navigation, wider availability of ScheduleStream code and performance on commodity edge AI hardware, and quantitative comparisons of grasp controllers like Grasp-MPC against established benchmarks. ## Scoring Rationale The story reports multiple peer-reviewed research outputs that validate simulation-trained policies on physical robots with measurable success rates and speedups, which is notably useful for robotics practitioners. It is important but not a paradigm shift, so it rates as a solid, notable development. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)