Gradient-Based Planning Is Finally Getting Serious, and It's About Time

I called my old colleague Frank who's still doing safety work at a Tier 1 supplier. His reaction was basically "show me the compute requirements." Fair enough. The paper claims GPU-batched computation, which is promising, but industrial safety systems tend to run on much more conservative hardware. That gap between "works on an A100" and "works on a safety PLC" is, well, it's significant.

The Planning Piece

The second paper, Dream-MPC, tackles a different problem that's been bugging me since I first saw model-based RL demos at IROS years ago. Gradient-based planning with learned models has always underperformed gradient-free methods in practice. Which is weird, right? You'd think having gradient information would help. But the loss landscapes from learned dynamics models are apparently nasty enough that gradient descent just gets stuck.

Their solution is clever, actually, let me be precise, it's a combination of several clever tricks. They roll out a policy to get candidate trajectories, then refine each one with gradient ascent. They add uncertainty regularization (so the optimizer doesn't exploit model errors) and they reuse optimized actions across timesteps. Tested on 24 continuous control tasks.

I'll be honest, I'm more skeptical of this one for industrial deployment. The learned world model is still a black box, and "uncertainty regularization" is doing a lot of heavy lifting in that sentence. But as a research direction, it suggests that hybrid approaches (combining learned models with classical MPC structure) might be the path forward.

What This Means for the Floor

Neither of these papers is going to show up in a production cell next year. Maybe not in five years. The gap between "demonstrated on quadrotor hardware" and "certified for automotive assembly" is vast, and I've watched too many promising academic results die in that gap.

But something's shifting. The reachability work in particular addresses a complaint I've been making since at least 2018: that neural network controllers were basically untestable by traditional methods. If you can compute guaranteed bounds on system behavior, even conservative ones, you can start having real conversations with safety engineers.

The compute requirements remain unclear from the papers. That's going to matter. And I only found these two sources covering this specific angle, so I might be missing context from the broader ML safety community.

Still. 72 dimensions, real-time, with certificates. When I started in this industry, that would've sounded like science fiction. Now it's a preprint on arXiv.

Progress is slow until it isn't.

Robert "Bob" Macintosh · 5 hours ago · 3 min