Robots Are Finally Learning to Handle Soft Things. Here's Why That's Harder Than It Sounds.

The team's approach uses something called robust model predictive control, or MPC, which is a well-established control framework, but the novelty here is running it in parallel on a GPU and coupling it with what they call conformal prediction, a statistical technique for calibrating uncertainty bounds. In plain English, the system is trying to be honest about what it doesn't know about where the cloth is and how it'll behave, and then plan conservatively enough that it stays safe despite that uncertainty. They test it on obstacle avoidance, rope routing, cloth folding, and cloth smoothing, and report that it beats baselines on safety, speed, and task success.

Millisecond-speed planning is the headline number here, and it's a meaningful one. Previous approaches to this kind of problem often required planning times that made real-time control impractical. Whether those millisecond speeds hold up outside of the specific hardware and simulation configurations they tested is, honestly, unclear. I only found the one source on the implementation details, and the paper is brand new.

The second paper, describing a framework called Perceive-Simulate-Imitate (or PSI, which is a better acronym than most), attacks a different piece of the puzzle. The question it's asking is: can a robot learn manipulation skills by watching videos of humans doing tasks, without needing any robot-collected data at all? The answer is sort of yes, with an important asterisk.

Human videos are actually pretty good for teaching the post-grasp part of a manipulation task, the bit where you've already got the object in your hand and you're doing something with it. They're much less useful for teaching grasping itself, because robot hands don't look like human hands, and the geometry is just different enough that you can't copy directly. PSI's solution is modular: use a dedicated grasp generator for the grasping part, and use the human video data for the downstream motion. The clever bit is a simulation filtering step that labels which grasps are actually compatible with the task the robot needs to perform afterward. A stable grasp isn't always a useful grasp, and that distinction matters a lot in practice.

So what

Here's my take, for whatever it's worth. Both of these papers are addressing real bottlenecks, not invented ones. The deformable manipulation problem has been a genuine obstacle to deploying robots in environments that involve fabric, cables, or any material that doesn't hold a fixed shape. That covers a huge range of practical applications, from warehouse logistics to surgical robotics to, yes, eventually household tasks.

The learning-from-video angle is interesting because it's trying to solve a data problem. Robot learning is hungry for training data, and collecting that data with actual robots is expensive and slow. If you can bootstrap from the enormous amount of human video that exists in the world, that's potentially a significant shortcut. The PSI paper is careful to note its limitations, specifically that it tested on prehensile manipulation tasks and that the simulation filtering step adds its own complexity. It's too early to say how well this generalizes beyond the specific tasks they evaluated.

What I keep coming back to is the gap between academic benchmarks and real-world deployment. Both papers do include hardware experiments, which is better than pure simulation, but the tasks are still controlled, the environments are still relatively structured, and the objects are still relatively well-behaved examples of deformable things. Real cloth in a real laundry situation is messier, wetter, heavier, and more unpredictable than a lab towel on a clean table. Real cables in a real wiring harness are tangled in ways that no simulator has fully captured yet.

Call me old-fashioned, but I think the honest version of this story is: these are genuinely good results that advance the state of the art in two specific and important subproblems, and we're probably still five to ten years away from seeing them translated into products that work reliably outside a lab. Maybe less if the GPU-parallel simulation approach turns out to scale as well as the authors suggest. Maybe more if the real-world robustness problems turn out to be harder than the benchmarks imply.

Some researchers in this space argue that the simulation-to-real gap is nearly closed for rigid objects and closing fast for deformable ones. Others counter that we've been saying that for a decade and the gap keeps revealing new layers. This raises questions about... well, multiple things, including whether our benchmarks are actually measuring the right things, and whether the field has the right incentives to be honest about failure cases.

What happens next

The CORD-SLS paper's GPU-parallel approach is worth watching, because it suggests a path toward scaling this kind of robust control to more complex scenarios without sacrificing the real-time performance that makes it practical. If the approach holds up under more adversarial conditions, and if the conformal prediction calibration works as well on novel objects as it does on the test set, that's a meaningful step.

The PSI framework's value proposition is essentially about data efficiency. If you can train a useful manipulation policy from human video without collecting robot data, you dramatically lower the cost of teaching robots new tasks. The question is how far that scales and how much the simulation filtering step can be automated. Right now it still requires a fair amount of careful engineering.

Both papers are from academic groups, and the companies paying attention to this space, the robot arm manufacturers, the warehouse automation players, the surgical robotics startups, will be looking at whether these techniques can be productized. That's a different problem than publishing a paper, and it's where most good academic robotics research goes to have a long, quiet conversation with reality.

I've been covering tech long enough to know that the papers that actually change the industry aren't always the ones that get the most attention when they drop. Sometimes it's the careful, methodical work on hard subproblems that ends up mattering most, because it removes the bottlenecks that were blocking everything else. Deformable manipulation has been one of those bottlenecks for a long time. Whether these specific approaches are the ones that finally crack it open, I genuinely don't know yet. But the direction is right, and the work is serious. That's more than you can say for a lot of what crosses my desk.

Aisha Patel · 10 hours ago · 10 min