NVIDIA Research is presenting 28 papers at ICRA 2026 with 8 specifically targeting simulation-to-real transfer, the bottleneck that has kept embodied AI in demo state. Concrete numbers from the named papers: COMPASS shows 4.5× improvement in average success rate vs imitation learning with ~80% on real robots via residual RL in Isaac Lab with no real-world data. Grasp-MPC reports 75% overall success on real robots vs 41% baseline, trained on 2 million simulated trajectories across 8,000 objects using cuRobo and GraspGen. PEEK reports 41× real-world accuracy improvement on sim-only policies and 2-3.5× gains for VLA models, via vision-language model guidance at image level.
The stack story under these numbers is what matters for builders considering embodied AI. NVIDIA's Isaac Lab is the simulation environment; cuRobo handles motion planning; GraspGen provides grasping datasets; Jetson runs on-robot inference. The 8 papers are bringing this stack from "expensive PhD project" to "industrial process" — multi-arm pharma coordination at 3× speedup (ScheduleStream on Jetson), precise assembly with 38% success-rate improvement and 30% cycle-time reduction (SPARR), multi-step assembly at 91% simulation success and ~11% improvement over baselines (Refinery), runtime candidate action verification with up to 15% gains (SEAL), and zero-shot transfer to real tree branches via synthetic trees generated from biological growth equations (Deformable Cluster Manipulation). The training compute is non-trivial (2M trajectories × 8K objects) but the resulting policies transfer without real-world data collection, which is the actual cost-saver.
The ecosystem read for builders: the "robots in simulation are easy, robots in reality are hard" gap is closing, and the methodology is converging on a common stack. Domain randomization remains the foundation, but the field is layering on residual policy learning (COMPASS), real-time motion correction (SPARR, Grasp-MPC), and VLM-guided perception (PEEK). The NVIDIA stack is the de facto reference implementation because the components are open or available, not because of vendor lock-in. The honest caveats: the 41× PEEK number is on sim-only policies that were near-zero in the real world, so the absolute starting point matters; most baselines are NVIDIA's own internal numbers from prior work, not head-to-head against other robotics frameworks; and the papers are blog-summarized, not yet through peer review at the time of writing. Worth tracking which numbers survive ICRA discussion.
If you build robotics applications Monday morning: the sim-to-real recipe is now reproducible enough that domain-randomization-plus-VLM-guidance is a default starting point, not a research direction. If you fund robotics startups: the cost curve for getting a manipulation or grasping policy from sim to deployable shrunk this year — the 75% real-world grasping number on novel objects in clutter is the practical milestone to mark.