Humanoid Robots Are Learning to Kick Footballs and Play Air Hockey. That's More Important Than It Sounds.

For context, that's not Cristiano Ronaldo. But it's not nothing either.

The simulation numbers are even more dramatic. Free-kick shot error 48.6% lower than prior baselines. Shoot velocity 2.96 times higher. Those are big gaps.

Why the curriculum approach matters.

You might be wondering why this is interesting beyond the specific soccer application, and that's the right question to ask.

The tension RoboNaldo is resolving, between rigid reference tracking and unconstrained reward-driven exploration, shows up everywhere in robotics. Motion tracking gives you stability but no flexibility. Pure task reward gives you flexibility but the robot basically has to rediscover physics from scratch. The curriculum approach threads that needle by using the human reference as a starting point and then systematically relaxing it.

This is sort of a general-purpose insight about how to train humanoids for high-impulse physical tasks. Kicking a ball is one instance. Pushing a heavy object, catching something thrown at you, getting up off the floor quickly, these all involve the same fundamental problem. Brief, forceful, destabilizing contacts that the robot has to handle without toppling.

I initially thought the soccer framing was just marketing. After reading more carefully, I think the soccer task was chosen precisely because it's a clean, measurable instance of a hard general problem. The target error metric gives you something concrete to optimize and evaluate. That's actually good experimental design.

Meanwhile, manipulation is getting smarter too.

The second paper, on arXiv about Interaction-weighted Resampling, tackles a different but related problem: how do you get a robot to learn object manipulation from scratch when the physics keeps changing in discontinuous ways?

Contrastive reinforcement learning, or CRL, has been doing well in locomotion and simpler control tasks. The idea is to learn structured representations of how states relate to each other, which helps with planning. But it's struggled with manipulation, and this paper has a specific diagnosis for why.

When a robot contacts an object, the underlying physics changes discontinuously. Before contact, the robot arm and the object are separate systems. During contact, they're coupled. After contact, they're separate again but with different momenta. The paper formalizes this as a piecewise-smooth Markov process, and argues that standard CRL energy functions just aren't built to represent these mode boundaries well.

The fix, Interaction-weighted Resampling (IWR), is conceptually clean. During training, the method resamples more densely around those contact transition phases, before, during, and after interactions, so the learned representation actually captures the mode boundaries that matter for planning.

Across a range of tasks including 2D dynamic control, robotic manipulation, and robot air hockey, IWR shows a 19.8% average improvement over prior CRL methods in simulation. The real-world air hockey result is the headline: the first goal-conditioned robot air hockey agent capable of hitting specified goals, improving success rate from 25% to 60%.

Tbh, going from 25% to 60% success on a task like that is a meaningful jump. Air hockey involves fast, unpredictable puck movement and requires precise, well-timed hits. It's another instance of that same general problem: high-speed contact with objects whose physics changes on contact.

Two papers, one underlying theme.

What I find genuinely interesting here is that both papers are circling the same core challenge from different angles. Contact-rich, high-impulse physical interaction is where current robot learning methods tend to break down. The dynamics are discontinuous, the timing windows are tight, and the consequences of getting it wrong (missed kick, knocked-over robot, puck flying the wrong direction) are immediate and obvious.

RoboNaldo addresses this through curriculum design at the motion level. IWR addresses it through representation learning at the dynamics level. They're not the same approach, and they're not directly comparable, but they're both trying to give robots better tools for handling the moments when physics gets complicated.

It remains unclear how well either of these generalises beyond their specific test domains. The RoboNaldo paper tests on a Unitree G1 in controlled conditions. The IWR real-world demo is a single robotic arm on an air hockey table. Both are limited settings, and I'd want to see these methods stress-tested on messier, less structured tasks before drawing strong conclusions about generalisability. That's not a criticism so much as the obvious next question.

What this means for humanoids specifically.

I cover humanoids, so I'll be direct about why I think the kicking paper matters for that space in particular.

Humanoid robots are increasingly being deployed in environments where they'll need to interact physically with objects that push back. Moving boxes, opening doors, navigating crowds. These aren't gentle tasks. The robot will get bumped. It will have to apply force. It will need to recover from contact without falling over.

The whole-body stability problem that RoboNaldo is solving for soccer is the same problem a warehouse humanoid faces when it's pulling a stuck drawer or a domestic robot faces when it's helping someone stand up. The specific motion is different. The underlying challenge is not.

I should be honest that I'm extrapolating here. The paper doesn't claim to solve humanoid manipulation broadly. It's a soccer shooting paper. But the curriculum framework and the stability results feel like they have legs beyond the football pitch, and I think the researchers know that.

The honest bottom line.

Honestly, I'm not sure either of these papers is a turning point on its own. Progress in robot learning is usually incremental, and it's easy to overread individual results. A 60% success rate on air hockey is good; it's not deployment-ready. A 0.73 metre average shooting error from 3 metres is impressive for a humanoid; it's not going to trouble a professional goalkeeper.

But the direction of travel matters. Both papers are pushing into territory where robots have historically been weak: fast, contact-heavy, physically demanding tasks where the physics is unforgiving. The fact that two separate groups, using different methods, are making measurable progress on this class of problem in the same week suggests the field is genuinely moving.

Sports demos aren't just demos. Sometimes they're the hardest version of the problem you actually care about.

The chip giant's latest numbers look like an AI infrastructure story. But if you're watching humanoids, there's something more interesting buried in there.

Sarah Williams · 10 hours ago · 4 min