MPC Is Getting a Neural Makeover, and It's About Time

Model Predictive Control has been the gold standard for robot motion planning for decades now, and it's also been a computational nightmare for just as long. A batch of new papers out this month suggests researchers are finally getting serious about fixing that.

I'll be honest, when I was at Kuka we had MPC running on some of our more advanced cells, and the joke was always that by the time the controller figured out the optimal trajectory, the part had already moved. That's an exaggeration, but only just. The computational burden of solving those optimization problems in real-time has kept MPC out of a lot of applications where it would otherwise shine.

The Neural Network Shortcut

The most straightforward approach comes from a team working on 3-DOF manipulators, detailed in a new arXiv paper. Their idea is simple enough: use behavior cloning to train a neural network that mimics what the MPC would do, but faster. They're claiming a 3x reduction in inference latency with an 84.98% success rate under what they call "relaxed tolerances."

Now, that success rate number caught my eye. 85% sounds good until you're running a production line and that other 15% means crashed robots or scrapped parts. The authors are upfront about this, they note there's a "precision gap under strict tolerances" and that the neural approximation captures the global trajectory but struggles with terminal steady-state error. In plain English: it gets you most of the way there but fumbles the landing.

This is exactly the sort of trade-off that keeps engineers up at night. Do you take the speed and accept the occasional miss? Depends entirely on your application. Pick-and-place in a warehouse? Maybe. Surgical robotics? Absolutely not.

What's interesting is their finding that static architectures outperformed temporal variants like RNNs. The instantaneous state observation was enough. I called my old colleague Hans at Siemens about this, and he wasn't surprised. "The robot doesn't need to remember where it was," he said. "It needs to know where it is and where it's going." Fair point.

• Behavior cloning can approximate MPC with significant speed gains
• Success rates drop under tight precision requirements
• Static neural networks outperform recurrent architectures for this task
• Terminal accuracy remains a problem that needs more work
• Trade-off is application-dependent, no free lunch here

Separately, there's work on what researchers are calling "world models" for manipulation, which is basically teaching robots to predict what happens when they do things. A comprehensive catalogs the current landscape, and it's fragmented. You've got latent dynamics models, video generators, physics simulators, and various hybrid approaches all competing for attention.