NVIDIA Research demonstrated significant advances in bridging the gap between simulation and real-world robotics at the International Conference on Robotics and Automation. Eight accepted papers highlight how the company's simulation-to-real transfer methods are enabling robots to perceive, reason, and act reliably in dynamic, unpredictable environments. These innovations address critical challenges across the robotics stack, including multi-arm coordination, generalized navigation, precise grasping, and complex assembly. In the domain of multi-arm coordination, the new ScheduleStream framework utilizes GPU computing to allow multiple robotic arms to plan and operate in parallel rather than sequentially. Deployed on NVIDIA Jetson edge AI platforms, this approach achieved a threefold speedup in planning scenarios, with code available for public access. For navigation, the COMPASS policy framework enables robots to adapt their movement strategies across different body types without real-world training data. By combining imitation learning with residual reinforcement learning in the Isaac Lab simulator, COMPASS improved average success rates by 4.5 times over baselines and achieved an 80% success rate across 20 real-world trials on autonomous mobile robots and humanoids. Grasping capabilities have also been enhanced through Grasp-MPC, which adaptively corrects motion in the final stages of an approach, mimicking human tactile feedback. Trained on millions of simulated trajectories, the system reached a 75% success rate on real robots handling novel objects in clutter, a substantial improvement over the 41% baseline. Additionally, Deformable Cluster Manipulation enables robots to handle tangled materials like branches or cables as single units. Using biological growth equations to generate synthetic training data, the system successfully deployed in a zero-shot manner to clear physical branch clusters. Precision assembly remains a difficult hurdle due to physical imperfections ignored by simulators. The SPARR method addresses this by splitting the task: a simulated policy learns the general strategy, while an on-hardware layer learns to correct discrepancies using camera data without human intervention. This resulted in a 38% increase in success rates and a 30% reduction in cycle time. For multi-step assembly sequences, the Refinery framework learns how to position components to facilitate subsequent steps, achieving 91% simulation success and notable real-world improvements. Cognitive and planning layers were further strengthened by the PEEK pipeline, which uses vision-language models to help robots filter out visual noise and focus on relevant objects. This innovation delivered a 41-fold accuracy improvement for policies trained purely in simulation. Furthermore, the SEAL method, developed in collaboration with Carnegie Mellon University and others, ensures robots execute actions that match their verbal instructions by evaluating candidate action sequences at runtime. This reduced execution errors by up to 15% across various conditions. Beyond specific algorithms, NVIDIA is expanding research infrastructure with large open datasets, including the Physical AI Dataset, which has surpassed 15 million downloads, and the Isaac GR00T X Embodiment Sim. Academic institutions such as MIT, ETH Zurich, and the University of Texas at Austin are increasingly leveraging these tools, with nearly 50 recent papers citing NVIDIA-accelerated simulation and robotics learning. These developments collectively signal a shift toward generalizable, reliable embodied autonomy outside the laboratory. Developers are encouraged to utilize the Isaac Lab and Isaac Sim environments to accelerate their own robotics projects.