A new simulator lets a million virtual robots argue over the same slice of 5G spectrum, all without leaving a single GPU.
Researchers built Isaac-Net, a GPU-batched 5G New Radio module that plugs into Nvidia's Isaac Lab physics engine. Instead of modeling network delay as a random number tacked onto each message, it simulates every 0.5-millisecond scheduling slot for thousands of parallel robot environments at once, matching how an actual 5G base station decides who gets to transmit. Tested against the established ns-3 5G-LENA simulator, its delay estimates land within 5 to 10 percent low on networks it was not tuned for, and about 9 percent high when 32 robots per environment compete for bandwidth. The payoff is scale: Isaac-Net keeps a full network simulation running for roughly one million robots on a single GPU, at 83 percent of the speed of running physics alone.
The detail that matters most is Age of Information, essentially how stale a robot's last network update is. The simpler approach most GPU simulators use today, giving every message an independent random delay, underestimates that staleness by about three times in the tail. Robots trained against that simplified network would learn to trust data that is, in practice, older and less reliable than their training ever showed them.
That is a meaningful gap for anyone coordinating warehouse or factory robots over shared 5G, and it is also why the team is open-sourcing the code rather than just the paper.