Network-in-the-Loop at Scale: GPU-Batched 5G Simulation for Massively Parallel Robot Learning

Can a GPU-batched 5G simulator keep pace with massively parallel robot learning while faithfully reproducing the delay and Age of Information behavior of packet-level network simulators like ns-3 5G-LENA?

Massively parallel GPU robot simulators train multi-robot policies in thousands of environments, but they replace the network with an independent delay per message, ignoring how each robot's traffic affects its teammates. Packet-level simulators such as ns-3 5G-LENA capture this contention but run one scenario per CPU process and are far too slow to match thousands of parallel training environments.

Before Isaac-Net, network-in-the-loop training either used independent per-message delays inside GPU robot simulators or packet-level simulators such as ns-3 5G-LENA running one scenario per CPU process. The former misses queueing and contention effects, and the latter cannot scale to thousands of environments.

Isaac-Net is a GPU-batched 5G New Radio module that advances the uplink for thousands of environments in lockstep with Isaac Lab physics. It simulates every 0.5 ms scheduling slot, in which the base station decides which robots transmit, for all environments simultaneously, so network delay depends on teammate traffic.

Across experiments, Isaac-Net's NR engine reproduced ns-3 5G-LENA median delay, with median delay 5 to 10 percent low on an unseen carrier and 9 percent high at 32 robots per environment in closed loop. It also reproduced Age of Information, while an independent per-message delay left the Age of Information tail different from the reference; exact tail metrics are not given in the abstract.

Not reported in the abstract. Evident limitations include simulation-only evaluation, no stated real-robot or fleet trials, no mention of code release, and unstated GPU compute overhead.