The Old Guard Meets the New: MPC Is Getting a Machine Learning Makeover

The key points from recent MPC-learning hybrid work:

• IFM framework combines conventional MPC with imitation learning, then fine-tunes with limited RL exploration
• EXACT-MPPI handles complex robot footprints without the usual convex approximations that throw away feasible motions
• SM2ITH (terrible acronym, by the way) adds human motion prediction to mobile manipulation tasks
• SDPG claims to train visual control policies in a few hours on consumer GPUs

I called my old colleague at Siemens last week, and he mentioned they're seeing similar trends in their research partnerships. The pure learning crowd and the pure control theory crowd are, well, finally talking to each other. It's about time.

The EXACT-MPPI work from another recent paper addresses something that's driven me crazy for years. Real robots aren't circles or rectangles. They've got arms sticking out, they're carrying loads, they're shaped like, honestly, like weird lumpy things. Most planners just inflate the footprint until it's a conservative blob, which means the robot thinks it can't fit through gaps it absolutely could fit through.

These researchers are doing exact signed-distance calculations for arbitrary footprints, running the whole thing on GPU. No training required, which means you can change the robot's shape (bolt on a new tool, pick up a pallet) and the planner just adapts. That's practical engineering, not just paper results.

What remains unclear is how well any of this transfers to truly unstructured industrial environments. The papers show impressive simulation results and some hardware demos, but simulation and a controlled lab are very different from a warehouse floor where someone's left a pallet in the wrong place and there's oil on the concrete. I've seen too many demos that worked perfectly until they didn't.

The human interaction piece is interesting too. The SM2ITH framework from researchers working on mobile manipulation tries to predict how humans will react to the robot's movements. They tested on the Stretch 3 and a Ridgeback-UR10, including scenarios with "adversarial human behavior," which I assume means someone deliberately getting in the robot's way. That's actually useful testing, because in my experience, warehouse workers will absolutely mess with robots if they're bored enough.

I'll be honest, I'm cautiously optimistic about this direction. The pure learning approaches always felt like they were throwing away decades of good engineering for the sake of end-to-end elegance. But the pure model-based approaches (and I built plenty of them) always hit a wall when the real world got messy. Maybe the hybrid path is the right one.

Or maybe I'm just getting sentimental in semi-retirement. Either way, it's more interesting than what I was reading five years ago.

Robert "Bob" Macintosh · Yesterday · 4 min