Robotics shift: simulation plus data

Robots are being trained in two places at once now: in virtual worlds built with physics software and in real warehouses labeled by humans. (reuters.com, cvat.ai) Cadence Design Systems and Nvidia said on April 15 that they are combining Cadence physics engines with Nvidia artificial intelligence models that train robots inside computer simulations. The companies said the goal is to cut the time it takes to teach robots useful tasks. (reuters.com) Cadence described the broader tie-up as an expansion announced at CadenceLIVE Silicon Valley 2026, with work spanning physics-based simulation, digital twins and physical artificial intelligence systems. Cadence said the collaboration will use Nvidia CUDA-X, artificial intelligence physics and Omniverse software, and that some engineering workflows could see speedups of up to 100 times. (cadence.com) Simulation is a video-game-like test range for robots: engineers can change materials, lighting, friction and sensor settings before a machine touches a real shelf or worker. Nvidia says Isaac Sim is built to develop, test, train and deploy robots in realistic virtual environments and supports synthetic data generation, reinforcement learning and digital twin simulation. (developer.nvidia.com, github.com) That still does not remove the need for labeled real-world data. A warehouse robot can collect camera images, LiDAR depth maps and motion data at the same time, and those streams have to be aligned and tagged so the model learns what an object is, where it is and how it is moving. (cvat.ai) In warehouses, the hard cases are ordinary ones: reflective wrap, moving forklifts, shifting pallets and workers crossing aisles. CVAT says even small labeling errors in three-dimensional point clouds can make a mobile robot stop abruptly or misjudge clearance. (cvat.ai) The mix of simulation and annotation is showing up as robot deployments keep rising. The International Federation of Robotics said 542,000 industrial robots were installed worldwide in 2024, the second-highest annual total on record, and more than double the level of a decade earlier. (ifr.org) That growth helps explain why suppliers are pitching larger annotation pipelines for robotics, especially for warehouse and logistics systems that depend on multimodal sensor data. Macgence says robotics annotation work now includes image, video, LiDAR and sensor labeling for navigation, picking and obstacle detection tasks. (macgence.com) The industry’s near-term bet is not simulation instead of data, or data instead of simulation. It is more simulated practice before deployment, and more carefully labeled real-world footage after it. (reuters.com, developer.nvidia.com, cvat.ai)