The Real Breakthrough in Surgical Robots Isn't Where You'd Expect

They achieved 11 out of 20 full task successes. That might sound underwhelming until you remember this is autonomous surgery inside a human eye. The failure cases matter enormously here, and I wish the paper had more detail on what went wrong in those 9 failures. But 55% autonomous success on a task that currently requires years of surgical training? That's not nothing.

Now compare this to the humanoid work. HERO is a system for getting humanoid robots to grasp everyday objects, mugs, apples, toys, on surfaces ranging from 43 to 92 centimeters high. The headline number is 2.44 centimeters of end-effector tracking error, which they claim is 5.5 times better than previous methods. And they tested it in real environments: offices, coffee shops, the kind of messy spaces that break most robot demos.

The approach is modular rather than end-to-end. They use large vision models for scene understanding (basically, figuring out what's in front of the robot and where), and then a separate learned policy for precise arm and hand control. The key innovation is what they call residual-aware tracking, a system that combines classical inverse kinematics with a neural network that predicts how the real robot will deviate from the commanded motion, then corrects for it.

I think this modular approach is underrated. The end-to-end learning paradigm has dominated robotics AI for years, the idea that you throw enough data at a neural network and it figures everything out. But these researchers are basically saying: look, we have really good classical tools for some parts of this problem. Let's use them. It's less elegant, maybe, but it seems to work better.

Here's what connects these two papers for me. Both are about closing the gap between simulation and reality, between commanded motion and actual motion, between where you think the robot is and where it actually is. In eye surgery, that gap is measured in fractions of a millimeter, and the stakes are someone's vision. In humanoid manipulation, it's measured in centimeters, and the stakes are (currently) whether a robot can hand you a coffee.

But the underlying challenge is identical. Robots are precise in theory. In practice, cables stretch, joints have backlash, sensors drift, and the world is messier than any simulator can capture. The real breakthrough isn't making robots move faster or stronger. It's making them aware of their own uncertainty and able to correct for it.

I should note some limitations here. The eye surgery work was done on artificial eye models, not humans. The humanoid work, while tested in real environments, still operates in relatively controlled scenarios (objects on tables, not buried in cluttered drawers). Neither paper discloses failure modes in detail, which is frustrating. And both rely on expert demonstrations, which means someone still has to show the robot what to do before it can do it autonomously.

Still, I keep coming back to that 0.686 millimeter number. We're building robots that can operate with sub-millimeter precision inside the human body, trained from watching experts, without explicit depth sensors. That feels like it should be bigger news than the latest humanoid doing a backflip.

Maybe I'm just getting older and more interested in medicine than spectacle. But I think there's something here about what we actually need robots to do versus what makes for a good demo. Humanoids grasping mugs is a step toward robots that can help in homes and workplaces. Surgical robots with autonomous precision are a step toward procedures that don't depend on the steadiness of a human hand.

Both matter. But tbh, if I had to pick which one I'd want working on me, I know which I'd choose.

Aisha Patel · 2 days ago · 8 min