85% Success Rates and Human-Mimicking Arms: Humanoid Manipulation Is Getting Serious

The first paper, arXiv, introduces something called HALOMI, which stands for Humanoid Loco-Manipulation with Active Perception from Human Demonstrations. It runs on a Unitree G1 humanoid, which is one of those full-body bipedal robots that looks vaguely unsettling when it walks. The HALOMI team extended an existing data collection tool called Universal Manipulation Interface to capture both what the robot sees from its head camera and what it sees from wrist-mounted cameras, along with the full trajectory of head and hand movements. Then they built what they call a "manifold-constrained controller" that plans motion in a learned latent space rather than directly in physical joint space. The idea is that by constraining the robot's plans to a manifold of physically plausible behaviors, you avoid the brittleness that comes from asking the robot to execute motions it was never trained on.

The results across five real-world tasks, including navigation, grasping, bimanual manipulation, and what they describe as dynamic behaviors like tossing, averaged 85% success on the three tasks they formally measured. The other two got qualitative treatment only, which is worth noting. Still, 85% on a full humanoid doing multi-step tasks is not a number you see every day.

The second paper, also out of arXiv, takes a different angle. DexSynRefine is about dexterous hand manipulation specifically, using human-object interaction data as a starting point. The key insight here is that they don't try to directly retarget human hand motions onto robot fingers, which almost never works cleanly. Instead they treat the human data as a "motion prior," a structured guess about what plausible motion looks like, and then use reinforcement learning to physically ground that guess in reality. They also add a component that infers missing contact dynamics from the robot's own proprioceptive history, meaning the robot uses what it feels to fill in gaps the human data didn't capture.

The improvement over naive kinematic retargeting is somewhere between 50 and 70 percentage points depending on the task. That's a big number. It's also the kind of number that makes me want to see an independent replication before I get too excited, but taken at face value it suggests the gap between "human does it" and "robot does it" is genuinely closing.

The Last Meter Problem Nobody Talks About

Here's the thing that doesn't get enough attention in mainstream robotics coverage. Even if a robot arm can grasp an object perfectly, it only gets the chance to try if the robot's base is positioned correctly in the first place. Getting close is easy. Getting close enough is hard. This is what roboticists call the "last meter" problem, and it's been a persistent bottleneck in mobile manipulation for years.

The third paper, from researchers at the RPM Lab at the University of Minnesota, takes this on directly. Their framework teaches a quadruped mobile manipulator to position itself precisely relative to a target object using only RGB camera images, no depth sensors, no LiDAR, no map. The robot takes in a goal image, its current camera views, and a text description of the target object, and figures out where to park itself.

What's interesting here is the generalization claim. They trained on a single instance of an object category and then tested on completely different instances of the same category, in new environments, under different lighting. The success rates were 74.58% on edge alignment (which checks whether the robot is oriented correctly using ground truth orientation data) and 89.42% on object alignment (which checks whether the robot is visually facing the target). Those numbers hold up across unseen objects and messy real-world conditions.

I'll be honest, I only found limited independent commentary on this specific paper so I'm working from the abstract and methodology description, and it's too early to say how these numbers hold up across a broader range of environments or object categories. The authors trained on a single object instance per category, which is impressively lean, but whether that scales to the chaos of a real warehouse or kitchen remains an open question.

So What

I've seen this movie before. A cluster of papers comes out, the numbers look good, the press gets excited, and then the hard part, actually deploying these systems in uncontrolled environments with real stakes, turns out to be much messier than the lab results suggested. I'm not saying that's what's happening here. I'm saying I've been burned enough times to read the fine print.

But there's something different about the current wave of humanoid manipulation research that I didn't see in, say, the self-driving car hype cycle of 2016 to 2019. The methodology is getting more honest. Researchers are publishing failure cases, reporting on tasks where they only have qualitative results, and being explicit about the limitations of their training data. HALOMI's team acknowledges the out-of-distribution brittleness problem directly. DexSynRefine is upfront that their approach depends on having decent HOI data to start with, which isn't always available. The Minnesota team flags that their evaluation metrics use ground truth orientation data in one case, which is a significant caveat for real deployment.

This is what mature engineering looks like, actually. Not the breathless press releases from companies promising full autonomy by next year (call me old-fashioned, but I stopped believing those around 2018). Real progress looks like incremental papers with honest error bars and explicit caveats.

The underlying technical convergence across all three papers is also worth noting. They're all, in different ways, trying to solve the same problem: how do you take cheap, abundant human demonstration data and make it actually useful for a robot that doesn't share your body plan? HALOMI does it with ego-view alignment and latent behavior manifolds. DexSynRefine does it with motion priors and residual RL. The Minnesota paper does it with language-conditioned segmentation and spatial reasoning. These aren't competing approaches so much as complementary ones, and the fact that multiple teams are converging on human demonstration data as the scalable path forward is itself a signal.

Whether any of this translates into a commercially viable humanoid robot doing useful work in the next two or three years, I genuinely don't know. The kids building these systems are smart and they're moving fast. But the gap between 85% in a lab and 99.9% in a hospital or factory is not a small gap. It's where most robotics companies have quietly gone to die.

Watch the deployment numbers, not the lab numbers. That's where the story actually gets interesting.

Mark Kowalski · Yesterday · 6 min