Video Friday: Do We Need Superhuman Humanoid Robots?

I spent the morning watching the latest footage from Unitree, and I have to be honest: my first thought wasn't 'the future is here.' My first thought was, 'who is going to pay for this?' The robotics world is currently obsessed with what I call the 'superhuman' trap. We are seeing machines that can backflip, sprint at terrifying speeds, and move with a fluid precision that makes a professional athlete look clumsy. It’s a technical marvel, but from a founder's perspective, it’s a red flag for over-engineering.

The Utility Gap

In the latest round of demos, we are seeing humanoid robots moving at speeds that would be dangerous for a human to stand near. Unitree is leading the charge here, showing off hardware that moves with a violent, jerky efficiency. The immediate question for anyone building in this space is: what is the commercial use case for a robot that can sprint to a second-floor window? Outside of high-speed package delivery in a war zone or incredibly dangerous industrial environments, most of what we are seeing is a flex of hardware engineering rather than a solution to a business problem.

As builders, we often fall in love with the 'cool' factor. We want to see how far we can push the motors and the actuators. But the reality of the market is that we don't need superhuman humanoids; we need reliable, boring machines that don't break. If a robot drops a glass in your kitchen, it doesn't matter if it can do a standing backflip. It matters if it can clean up the shards without scratching the floor or bricking its own OS. Right now, the gap between 'look what this can do in a lab' and 'this works in a house with a toddler' is still a canyon.

Learning Without Training: The GEN-1.5 Breakthrough

One of the more interesting developments lately isn't just how these robots move, but how they learn. A group of researchers recently showcased GEN-1.5, a foundation model for physical tasks. The claim is that it can learn a new skill in seconds from a single demonstration, without needing to be fine-tuned. This is 'in-context learning' for the physical world, similar to how we use LLMs today.

This is where the real value lies for founders. If we can move away from the 'training a model for six months to pick up a strawberry' phase and into a 'show the robot once and it gets it' phase, the economics of robotics change overnight. However, I’m still skeptical. The researchers admit these are short-horizon, simple tasks. And as a skeptic, I have to point out the 'pre-training' problem. When a model 'figures it out' on the first try, we have to ask if it really learned it on the fly, or if it just saw ten thousand similar examples in its training data that the developers forgot were there.

Retrofitting vs. Replacing

While the humanoid hype gets the views, the real money is moving into boring sectors like heavy construction. Companies like Gravis Robotics are taking a much smarter approach: they aren't trying to build a new robot from scratch. They are building retrofit kits. They take an existing, multi-ton hydraulic excavator and give it eyes (lidar and cameras) and a brain (onboard compute).

For builders, this is a masterclass in market entry. Instead of convincing a construction firm to buy a $500,000 humanoid to move dirt, you sell them a kit that makes their current fleet 20% more efficient. The robot works close to the physical limits of the machine, moving more material per hour than a tired human operator could. It’s not flashy, it’s not 'superhuman' in a way that looks good on TikTok, but it has a clear ROI. That is the difference between a science project and a business.

The Problem with Physical Generalization

We are also seeing a lot of progress in brachiation—robots swinging from bars like primates. The researchers at EVARL are using waypoint-guided reinforcement learning to get life-sized robots to navigate monkey bars. Again, the physics are impressive. Sim-to-real transfer is getting better every day. But as a founder, I look at that and think about maintenance costs. Every time a 200-pound robot swings its weight, the stress on those joints is astronomical.

We have to ask ourselves: are we building these forms because they are the most efficient way to solve a problem, or because we are obsessed with mimicking biology? A robot inspired by a Mexican jumping bean (like the BeanBot project) might be more useful for certain niche sensor deployments than a humanoid that requires constant calibration. We need to be careful not to let the humanoid form factor become a constraint on our creativity.

Takeaway for Builders

If you are building in the AI or robotics space right now, my advice is to ignore the 'superhuman' demos. They are useful for raising VC money, but they are distractions from the hard work of building utility. Focus on the 'one-shot' learning capabilities and the software layers that allow machines to understand their environment. The real winner of the robotics race won't be the one whose robot runs the fastest; it will be the one whose robot is the easiest to teach.

The goal isn't to build a machine that can replace a human athlete; it's to build a machine that can be trusted to do a task when no one is watching.

We are seeing the building blocks of general physical intelligence, but we are still missing the bridge to mass adoption. Stop trying to make your robots 'superhuman' and start trying to make them 'super-useful.' The former gets you a viral video; the latter gets you a company.

Read the original at IEEE Robotics →