Two Papers That Actually Matter for Industrial Motion Planning (And Why Nobody's Talking About Them)

Now, I called my old colleague at Siemens last week about something unrelated, and we ended up discussing this exact problem. His team still handles it the old-fashioned way: conservative speed limits and lots of manual path tweaking. The idea that you could have the planner automatically adapt its step size based on local kinematic conditioning... that's the kind of thing that could save integrators hundreds of hours per installation.

The second paper tackles a related but distinct problem: what happens when your robot has multiple objectives that conflict with each other. arXiv describes a method using nullspace projections to resolve these conflicts dynamically. The application they tested was pushing non-convex objects, which is notoriously difficult because the contact dynamics are unpredictable.

Their results are striking:

• 100% success rate across 100 configurations for their method
• 0% for steepest-descent baseline (basically, it gets stuck in local minima every time)
• Roughly 55% for diffusion policy, which requires demonstrations and training
• The same approach transferred directly to a real robot without retraining

That last point is what gets me excited. So much of the machine learning work I see requires massive datasets and breaks the moment you change the gripper or move to a different robot. This is old-school control theory done right, and it generalizes.

What This Means for the Factory Floor

I want to be careful not to oversell this. These are academic papers, not commercial products. The gap between "works in a lab" and "works in a production environment with dust, temperature swings, and operators who don't read manuals" is substantial. It remains unclear how these methods would handle the kind of sensor noise and calibration drift you see in real installations.

But the underlying ideas are sound, and they address problems that have been annoying robotics engineers since I started in this business. When I was at Kuka, we had customers who would literally redesign their entire cell layout because we couldn't get the arm to reach certain positions without joint limit issues. That's expensive. That's the kind of thing that makes companies decide robots aren't worth the hassle.

If these verified planning methods make it into commercial controllers (and I'd bet Fanuc and ABB are already looking at similar approaches internally), it could meaningfully reduce integration time and expand the range of applications where robots make economic sense.

The humanoid stuff is flashy. I get it. But the unsexy work of making industrial arms more reliable and easier to program? That's what actually moves the needle for adoption. I just wish more people were paying attention.

Everyone's excited about video world models and 4D representations, but having spent years actually deploying robots, I see some familiar patterns here.

Robert "Bob" Macintosh · 3 hours ago · 4 min