November is fully booked! Next classes blast off the week of Mon, Dec 7.

Reserve your spot
Humanoid Balance Control: A Robot Crossed the Monkey Bars
Table of Contents

Humanoid Balance Control: A Robot Crossed the Monkey Bars

A humanoid robot traversed monkey bars in September 2026. Here is what humanoid balance control actually requires, and why hanging is harder than walking.

On September 11, 2026, IEEE Spectrum’s weekly robot video roundup led with a humanoid robot jumping up onto a set of monkey bars and swinging across. It looks like a stunt. It isn’t. Humanoid balance control on a flat floor is a solved-ish problem that dozens of companies now ship; balance control while hanging from two thin steel bars, with your feet off the ground and your whole body swinging, is a different mathematical problem with a different failure mode. The gap between those two things explains most of what’s confusing about robot videos right now.

Key Takeaways

  • The robot came from ETH Zurich’s Robotic Systems Lab, and the published framing of the work is specific: “Traversing sparse 3D structures requires humanoid robots to perceive thin, overhanging geometry while executing agile, accurate whole-body motions.”
  • Walking balance keeps the robot’s center of mass over a support polygon on the ground. Hanging balance has no support polygon: the robot manages angular momentum around a grip point instead.
  • The perception half is as hard as the control half. Most legged robots build a 2.5D height map of the floor. A monkey bar is thin, has nothing under it, and breaks that representation.
  • This is a research demo, not a product. Evan Ackerman’s September 2025 IEEE Spectrum analysis found that even the leading humanoid companies had deployed “only a small handful of robots in carefully controlled pilot projects.”
  • For a kid, the useful lesson is the one engineers actually use: balance is not a feeling, it’s a prediction. The robot is constantly guessing where its body will be a quarter-second from now.

What actually happened on September 11, 2026

A humanoid robot traversed a monkey-bar structure by jumping onto it, moving across a sequence of sparse bar contacts, and landing safely. The demonstration came from the Robotic Systems Lab at ETH Zurich and appeared in IEEE Spectrum’s “Video Friday: Humanoid Robot Takes On Monkey Bars” on September 11, 2026.

Worth noticing what else was in that same roundup, because it maps the field well. Duke’s General Robotics Lab showed Humanoid v2. The Center for Robot-Assisted Search and Rescue revisited the robotics deployment at the World Trade Center. Unitree released UnifoLM-WLA-1.0, a mobile-manipulation foundation model the company describes as supporting cross-task and cross-end-effector generalization. The DRAGON Lab in Tokyo demonstrated an aerial manipulation robot using hybrid impedance-admittance control for contact-rich tasks. ETH Zurich also showed a quadruped doing parkour with an attention-based map encoder.

That’s one week. Five of those nine items are about contact: touching things on purpose, with force, without falling over. Contact is where robotics is currently hard.

Why hanging is a different problem from walking

Start with the definitional sentence, because the rest follows from it. Humanoid balance control is the continuous problem of keeping a robot’s body motion consistent with the forces its contacts can actually produce.

When a robot walks, its contacts are its feet, and the region bounded by those contact patches is called the support polygon. A standing robot stays up if the ground reaction force stays inside that polygon. Engineers formalize this in a few ways (the zero-moment point, the centroidal momentum, the capture point), but the intuition is simple: there is a patch of floor, and the robot’s weight has to push down through it.

Monkey bars delete the patch. The contacts are two hands wrapped around a cylinder maybe 3 centimeters across. There is nothing underneath the robot at all. Gravity is now pulling it down and the only thing resisting is grip force plus whatever torque the shoulders and wrists can generate. The system stops behaving like an inverted pendulum balanced on a point and starts behaving like a regular pendulum hanging from one, which is stable in position but violently unstable in timing. Release at the wrong moment and the body goes the wrong direction.

So the controller has to do three things it never does on a floor:

Manage angular momentum as the primary resource. A walking robot regulates momentum mostly by choosing where to step. A swinging robot regulates it by pumping the legs and torso, exactly as a child does on a swing, then has to spend that momentum precisely on the reach.

Treat grip as a force budget, not a binary. A foot on concrete can push with hundreds of newtons. A hand on a bar can only pull toward the bar, and only as hard as friction and finger strength allow. The planner has to stay inside that cone of achievable forces.

Commit before it can see the outcome. Once the robot releases one hand, the next contact either happens or it doesn’t. There’s no mid-air correction loop like the one a walking robot uses to adjust a footfall.

The perception half nobody mentions

Here’s a detail engineers would screenshot. Most legged robots do not carry a 3D model of the world. They carry a 2.5D elevation map: a grid where each cell stores one height value, the floor or the obstacle on it. That representation is cheap, fast, and good enough for stairs, rubble and hills.

A monkey bar breaks it completely. The bar has a height. Underneath the bar is empty air, which the same grid cell cannot also represent. And thin cylinders are among the worst targets for depth cameras and LiDAR: the laser or stereo match either hits the 3-centimeter bar or sails past it into the background, so the point cloud around a bar is sparse and noisy at exactly the resolution you need.

This is why the ETH framing names perception first. Takahiro Miki and colleagues, publishing in Science Robotics in January 2022, showed why: their perceptive locomotion controller for the ANYmal quadruped used an attention-based recurrent encoder to blend vision with proprioception precisely because “depth perception can degrade due to difficult lighting, dust, fog, reflective or transparent surfaces.” The robot learned to trust its own joint sensing when the camera lies. That controller completed an alpine hiking route inside the posted human time. The same principle, believe your body when your eyes fail, is what lets a robot commit to a bar it can only partly see.

The follow-on work is also instructive. In ANYmal Parkour (Hoeller, Rudin, Sako and Hutter, 2023), the ETH group trained a quadruped to chain jumping, climbing and crouching skills under a high-level navigation policy, reaching 2 meters per second across consecutive obstacles with no expert demonstrations. Hierarchy plus simulation plus a perception module that reconstructs obstacles. The monkey-bar result sits in that lineage.

How to Teach Your Kid About Humanoid Balance Control

The monkey bars are the lab. You already own them.

Ages 5–8: The freeze game

Have your kid stand on one foot, then slowly lean forward until they have to put the other foot down. Ask them to notice the exact moment they knew they were going to fall, it’s always earlier than the fall itself. That early moment is the robot’s job. Then do it with eyes closed. Most kids are dramatically worse, which makes the point about vision without any vocabulary at all. Finish with: stand on a pillow. Soft ground changes the force you can push with, and they’ll feel it.

Ages 9–12: Support polygon with chalk

Draw a shoe outline on the driveway, then draw the outline of both feet with a line connecting the outside edges. That connected shape is the support polygon. Have them lean until a part of their body is outside it. Now have them hold a backpack out to one side: the polygon didn’t move, but the center of mass did, and they’ll shuffle without being told to. Then hang from a bar and ask: where’s the polygon now? There isn’t one. That’s the whole article in thirty seconds.

Ages 13+: Measure the swing

Use a phone’s slow-motion camera to film them crossing monkey bars. Count frames between hand releases. Then have them try to cross without swinging: purely arm strength, body hanging still. Most kids can’t do more than two or three bars that way. Have them write down why: the swing is storing and releasing energy, and the release has to be timed to the top of the arc. That’s a controller. If they code, this is a natural bridge into writing a pendulum simulation and testing release timing numerically.

The question to ask: “If a robot could see perfectly but couldn’t feel its own arms, could it cross the monkey bars?”

Walking balance vs. hanging balance: what changes

FactorWalking on a floorHanging from bars
Contact geometryTwo foot patches forming a support polygonTwo grip points, no polygon
Force the contact can applyPush, large, in any direction inside the friction conePull toward the bar only, limited by grip
Primary control resourceWhere to place the next footAngular momentum and release timing
Recovery from errorAdjust the next footstep mid-strideMostly none once the hand leaves
World representation needed2.5D elevation map usually worksFull 3D occupancy, bars overhang
Sensor difficultyModerateHigh: thin cylinders are poor depth targets
Consequence of failureTrip, often recoverableFall from height, usually not
Commercial status in 2026Shipping in pilotsResearch only

What a parent should actually do with a video like this

Separate the demo from the deployment

A robot doing something impressive in a lab video is not a robot doing it in a warehouse. Charlie Kemp, cofounder and CTO of Hello Robot, put this as plainly as anyone in a September 28, 2026 profile: “It’s easy to watch videos of robots online and feel like robotics has been solved. But the truth is that there is a vast difference between a demo video and a robot being productive in the real world.” That sentence is worth reading aloud to a kid who just watched a robot backflip.

Check whether a human is driving

Many robot videos that look autonomous are teleoperated, partly or entirely. The monkey-bar run is research from a university lab, which usually means the policy is autonomous, but the default assumption for a company’s promotional video should be “I don’t know yet.” We wrote a fuller guide to spotting teleoperation in robot demo videos and to the human behind an “autonomous” home robot.

Ask what was cut

Research groups show the successful run. That’s normal and not dishonest: papers report success rates, videos report highlights. The honest question is how many attempts there were. If a kid internalizes “demos show the best take,” they’ve learned something transferable well beyond robotics.

Point at the boring robots too

The International Federation of Robotics reported on September 24, 2026 that the global operational stock of industrial robots reached 5 million units in 2025, up 9 percent, with more than 600,000 new installations in the year. Almost none of those look like humanoids. They’re arms bolted to floors, doing one thing, extremely well. That’s where the jobs are, and it’s a useful counterweight to a monkey-bar video. Our piece on what current robots can and can’t do goes into the comparison.

What not to do

Don’t tell a kid the video is fake. It usually isn’t, and the claim collapses the moment they find the paper. Teach calibration, not cynicism. “Real, narrow, and early” is the accurate frame, and it leaves room for the kid to be genuinely excited.

What to Watch For Over the Next 3 Months

  • Week 4: Look for a preprint or conference submission attached to the monkey-bar result. Research groups usually post the paper within weeks of the video, and the paper will report a success rate the video doesn’t.
  • Month 2 red flags: A company, not a university, announcing a humanoid that “climbs” or “traverses structures” in a product video, with no success rate, no paper, and no stated autonomy level. Also watch for the same footage recut with different music; that often signals there’s only one good take.
  • Month 3 self-check: Can your kid explain, unprompted, why a flat floor is easier for a robot than a ladder? If they reach for “because there’s nothing to balance on” rather than “because ladders are complicated,” the concept landed.

Frequently Asked Questions

Is this robot going to be sold?

No. This is university research from ETH Zurich’s Robotic Systems Lab, published through an IEEE Spectrum video roundup. Research platforms in this class cost in the hundreds of thousands of dollars and need a trained team. The techniques will show up in commercial robots over years, usually in simplified form.

Why do robots need to climb at all? Warehouses have floors.

Most don’t. The customers who care are inspection, search and rescue, and construction, where stairs, ladders and scaffolding are unavoidable. CRASAR’s work after the World Trade Center collapse, revisited in that same September 2026 roundup, is the founding case: the terrain was rubble, and wheels were useless.

My kid wants to build a robot that does this. Realistic?

Not the monkey bars. But the underlying idea scales down well. A servo-driven two-link pendulum that learns to swing itself up is a genuine high-school project, and it teaches the exact concept, momentum pumping and release timing. Start with a path from humanoid hype to building at home.

Does this mean humanoid robots are nearly ready?

The published evidence says no. In September 2025, IEEE Spectrum reported Bank of America Global Research projecting roughly 18,000 humanoid units shipped in 2025, against an installed base of about 500,000 industrial robots added in 2023 alone. Former Agility Robotics chief product officer Melonee Wise framed the bottleneck as demand, not capability: “I don’t think anyone has found an application for humanoids that would require several thousand robots per facility.”

What’s the single hardest part?

Reliability, not acrobatics. Wise noted that industrial customers expect around 99.99 percent uptime, roughly five minutes of downtime a month. A robot that crosses monkey bars once is a research result. A robot that crosses them ten thousand times without a failure is a product, and nobody has that.

Should I let my kid watch robot videos?

Yes, with one habit attached: after every video, ask “what would this robot fail at?” It converts passive watching into engineering practice, and kids are surprisingly good at it.


About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.


Sources

  1. IEEE Spectrum. (2026, September 11). “Video Friday: Humanoid Robot Takes On Monkey Bars.” IEEE Spectrum. https://spectrum.ieee.org/video-friday-disaster-response-robots
  2. Miki, T., Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., & Hutter, M. (2022). “Learning robust perceptive locomotion for quadrupedal robots in the wild.” Science Robotics, 7(62). https://arxiv.org/abs/2201.08117
  3. Hoeller, D., Rudin, N., Sako, D., & Hutter, M. (2023). “ANYmal Parkour: Learning Agile Navigation for Quadrupedal Robots.” https://arxiv.org/abs/2306.14874
  4. Ackerman, E. (2025, September 11). “Reality Is Ruining the Humanoid Robot Hype.” IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-scaling
  5. International Federation of Robotics. (2026, September 24). “Five Million Robots now Operate in Factories Globally.” IFR Press Release. https://ifr.org/ifr-press-releases/news/five-million-robots-now-operate-in-factories-globally
  6. Robots Guide. (2026, September 28). “A Day in the Life of a Roboticist: Charlie Kemp.” https://robotsguide.com/learn/a-day-in-the-life-of-a-roboticist-charlie-kemp
  7. Occupational Safety and Health Administration. “Industrial Robot Systems and Industrial Robot System Safety.” OSHA Technical Manual, Section 4, Chapter 4. https://www.osha.gov/otm/section-4-safety-hazards/chapter-4
Ricky Flores
Written by Ricky Flores

Founder of HiWave Makers and electrical engineer with 15+ years working on projects with Apple, Samsung, Texas Instruments, and other Fortune 500 companies. He writes about how kids learn to build, think, and create in a tech-driven world.