Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
A pair of research papers dropped recently that caught my eye, both dealing with what the academics call "whole-body manipulation." That's a fancy way of saying robots that push stuff around using their entire bodies rather than just a gripper on the end of an arm.
Look, I know this doesn't sound as exciting as the latest humanoid doing backflips. But after 12 years watching industrial automation evolve, I can tell you that the unsexy problems are usually the ones worth paying attention to.
Here's the thing about manipulation in robotics: we've gotten pretty good at pick-and-place. A Fanuc arm can grab a widget off a conveyor and put it in a box faster than you can blink. But that assumes the widget is light, the gripper fits it, and everything's in a nice controlled environment.
Real world doesn't work that way. Sometimes you need to move a 50kg crate across a warehouse floor. Sometimes the object is awkward shaped. Sometimes, and this was a constant headache when I was at Kuka, you just don't have the payload capacity for a gripper that can handle everything you need to move.
Pushing is what animals do when they can't carry something. Ants push food. Dogs nose their bowls around. It's a non-prehensile manipulation strategy (there's your ten-dollar phrase for the day) that sidesteps the whole grasping problem entirely.
Two new papers show reinforcement learning works better when we stop pretending AI can figure everything out alone.
Mark Kowalski · 33 mins ago · 6 min
One uses formal verification to know when exploration is safe; the other asks whether we can skip the safety oracle entirely. Both are wrestling with the same fundamental problem.
Aisha Patel · 2 hours ago · 8 min
Millimeter-accurate fisheye depth data and a clever low-light navigation hack both point to the same uncomfortable truth: we've been training robots on the wrong data.
James Chen · 4 hours ago · 4 min
Everyone's chasing humanoid demos, but two new perception papers solve problems that have quietly blocked real-world deployment for years.
The first paper comes from a team working on hexapod robots, published on arXiv. They've developed something called HeLoM, a hierarchical learning framework that lets a six-legged robot push objects using coordinated multi-limb control.
The approach is inspired by how insects work together, using multiple contact points and all those degrees of freedom to maintain balance while shoving things around. The high-level planner figures out the pushing behaviour, while a low-level controller keeps the robot from falling over. They tested it pushing objects of varying sizes and unknown physical properties to target locations.
The second paper, also on arXiv, goes in a completely different direction. It's a multilink multirotor called DELTA that can fly, roll on the ground, and manipulate objects. Each link has its own thruster, which lets the thing transform its shape for different tasks.
I'll be honest, the DELTA robot is the weirder of the two. It's trying to solve air-ground hybrid locomotion AND manipulation, which is ambitious to the point where I'm not sure what the practical application is yet. But the underlying control problem, maintaining stability while your robot is doing multiple things at once, that's genuinely hard.
This is where I have to hedge. The hexapod work includes real-world experiments, which is encouraging. They claim stable pushing of objects with unknown physical properties, and the videos (if you dig them up) look reasonably convincing.
The DELTA paper is more prototype-stage. They've demonstrated the motions, but it remains unclear how robust this is outside controlled conditions. The paper itself admits this is "the first demonstration" of this type of hybrid locomotion and manipulation, which in academic speak means "we got it working once and we're excited about it."
I called my old colleague at Siemens last week about something unrelated, and we ended up talking about legged robots in logistics. His view, which I sort of share, is that the control theory is advancing faster than the mechanical reliability. You can make a hexapod push a box across a lab floor. Making it do that 10,000 times without breaking is a different problem entirely.
When I started at Kuka in the early 2000s, mobile robots in warehouses were basically glorified forklifts on rails. The AGVs we worked with needed magnetic tape on the floor and couldn't handle anything unexpected.
Now you've got AMRs navigating dynamically, collaborative arms working alongside humans, and apparently hexapods learning to shove boxes around using reinforcement learning. The trajectory is clear even if the timeline isn't.
The hexapod approach is interesting because it potentially solves the "last ten feet" problem in warehouse automation. You can get pallets to a general area easily enough. Getting individual items from that pallet to exactly where they need to be, especially heavy or awkward ones, that's still largely manual labour.
Whether a six-legged pushing robot is the answer, I genuinely don't know. The form factor seems impractical for most facilities I've worked in. But the control methods, the hierarchical planning, the whole-body coordination, those could transfer to other platforms.
Two things. First, manipulation research is finally getting serious about contact-rich tasks that don't fit the pick-and-place paradigm. That's been a gap for years.
Second, and this is the part that fires me up a bit, we're seeing academic robotics tackle problems that industrial users have been complaining about forever. The disconnect between what researchers build and what factories need has always frustrated me. Papers like these, while still pretty far from deployment, at least feel pointed in a useful direction.
I'm not saying hexapod warehouse workers are coming next year. I'm saying the underlying capability, robots that can use their whole bodies to interact with heavy, awkward, unpredictable objects, that's worth watching. Even if the current prototypes look like something out of a sci-fi B-movie.