The Jitter Problem: Why Your Robot's Smooth Moves Still Look Like a Nervous Intern

The second paper takes a more radical position. The Neural Implicit Action Fields (NIAF) team argues that the whole approach of predicting discrete waypoints is "structurally misaligned with the intrinsic continuity of physical motion." They're not wrong. When I was at Kuka, we spent enormous effort on trajectory interpolation because the motion planners spit out waypoints and then you had to make the actual movement between them not look like a robot having a seizure. NIAF tries to skip that step entirely by outputting continuous action functions instead of discrete points.

The clever bit is using the vision-language model as what they call a "hierarchical spectral modulator" over a learned motion prior. In plain English: the high-level AI decides what the motion should generally look like, and a specialized motion network fills in the continuous details. They claim this enables "analytical differentiation" for velocity supervision, which should theoretically give you mathematically consistent acceleration profiles. Real-world tests showed stable impedance control, which is genuinely important if you're doing anything involving contact.

I called my old colleague at Siemens about this one, actually. His take was skeptical but interested. "Continuous representations are elegant," he said, "but I've seen elegant solutions die in noisy sensor environments." Fair point. The paper doesn't fully address how robust this is to the kind of perception noise you get in real facilities.

The third paper, Fisher-Preserving Guidance, attacks a related but distinct problem. When you try to guide a diffusion policy at test time (say, to avoid an obstacle that wasn't in training), you can push the outputs into regions the model was never trained on. The result is what they call "Fisher drift," basically the model starts hallucinating physically implausible motions because you've wandered off the manifold of sensible actions.

Their solution is training-free, which is appealing from a deployment standpoint. You don't want to retrain your policy every time the warehouse layout changes. They compute what they call a "Fisher-preserving update" using a low-rank Jacobian factorization, which, and I'm simplifying here, keeps the guidance corrections within the space of motions the model actually understands. They tested it on navigation tasks including real robots, which gives me slightly more confidence than pure simulation results.

Here's my concern with all three approaches. They're solving the motion quality problem at the policy output level, but in my experience, a lot of jitter comes from the control loop fighting with itself downstream. You can have a beautifully smooth trajectory command, and then the servo loops, the compliance controllers, and the safety systems all add their own noise back in. The NIAF paper at least acknowledges this by emphasizing impedance control compatibility, but it remains unclear how these methods interact with the full control stack on industrial hardware.

The other thing nobody's really addressing: what happens when the smooth motion is actually wrong? Sometimes that human demonstrator paused because they were double-checking alignment. Sometimes the jitter was corrective. We're assuming all high-frequency content is noise, but I've seen cases where it was signal. The FGO paper waves at this with talk of "preserving details necessary for successful task execution," but the mechanism for distinguishing good detail from bad noise seems to be, basically, trust the frequency bands we chose.

I'm cautiously optimistic about this direction of research. Motion quality has been a second-class citizen in the learning-based robotics world for too long, everyone focused on task success rates without caring if the robot looked like it was being operated by a caffeinated squirrel. But we're still in early days. These are benchmark results, mostly simulation, with limited real-world validation. I'd want to see how they hold up after six months of 24/7 operation in a dusty fulfillment center with flickering lights and pallets that aren't where the map says they should be.

For now, I'll keep watching. And probably keep complaining. Old habits.

Mark Kowalski · 22 hours ago · 5 min