Cadence and Nvidia expand partnership to train robots on more accurate simulations

Cadence and Nvidia are going after one of robotics' most painful problems: the gap between what a robot learns in simulation and how it behaves afterward in the real world. The companies announced the expanded partnership on April 15, 2026, at CadenceLIVE Silicon Valley in Santa Clara. The goal is simple to state and hard to execute: make robot training data realistic enough that physical AI reaches real-world deployment faster instead of getting stuck between the lab and the factory floor.

For Nvidia, this continues its bet on physical AI, a class of systems that must do more than generate text or images: they have to act in space, handle objects, move around, and account for material resistance, friction, collisions and mass. For Cadence, it is a logical step beyond its familiar image as a supplier of chip-design software. Besides EDA tools, the company has strong physics engines and multiphysics simulators that can model metal deformation, fluid flow, surface contact and other processes on which robotics depends heavily. The partners want to build exactly these models into the robot training loop.

The problem they are tackling is well known across the industry. Training a robot in the real world is slow, expensive and often unsafe: it takes hardware, space, engineers, protection against breakage and a huge number of repetitions. Simulation lets teams run scenarios faster and more cheaply, but only as long as the virtual environment reflects real-world physics accurately enough.

If an object slides differently in simulation than in reality, if a surface is too perfect, or if gripper contact is computed in a simplified way, the model will learn the wrong behavior. The robot that looked convincing on screen then starts making mistakes in a warehouse, workshop or assembly line. The more accurate the synthetic data, the more useful the training, and the whole logic of the deal rests on that.

Technically, the joint stack is meant to connect several layers into one working pipeline. On Nvidia's side, it includes the open Isaac libraries for robot simulation and training, the Cosmos world models, and Jetson hardware for running systems at the edge and inside the devices themselves. On Cadence's side, it includes high-fidelity multiphysics simulation and additional test environments such as VTD and VTDx for complex scenarios.

In the official description of the partnership, the companies describe a pipeline orchestrated by AI agents: these agents will coordinate the stages from data preparation and training orchestration to policy optimization, validation, field data and post-deployment feedback. In other words, this is not about a pretty 3D picture but a closed loop in which virtual training is constantly checked against the machine's real behavior.

More broadly, it is also a signal of how the AI infrastructure market is changing. Nvidia has steadily built links between compute, simulation, digital twins and industrial software: it previously deepened partnerships with Siemens and Dassault Systèmes, and now it is strengthening its robotics push through Cadence. For Cadence, it is a chance to claim a place not only in the chip-development chain but also in the growing layer of tools for physical AI.

In the press release on the expanded partnership, the companies also discuss agentic AI and digital twins for AI factories; for some engineering tasks, Cadence promises workflow speedups of up to 100x, although in robotics the key factor remains not speed itself but simulation fidelity and confidence when moving systems into real environments.

The main takeaway: the robotics race depends less and less on the model alone and more on data quality, simulation physics and continuous feedback after deployment. If the Cadence-Nvidia combination really does narrow the sim-to-real gap, robot makers will be able to test new skills faster, break hardware less often in the early stages and ship systems into real tasks with more confidence. That does not mean the problem is solved for good: the real world is still messier, noisier and less predictable than any digital copy. But the partnership shows where the market is heading: toward more accurate simulation, tighter integration of software and hardware, and a more practical path from model training to a working robot.