Cristiano Roboto

Jeff’s Note: Before we get to today’s issue, I want to put something on your radar…

My colleague Larry Benedict and I are planning an exciting event next week called the Overnight AI Gains Summit where we’ll reveal a bold trading strategy for profiting from
AI.

They’re called “zero-day” trades – as in “zero days to expiration.” These trades are fast and have explosive potential… especially when applied to the right AI stocks.

When used properly, this strategy can potentially double your money or more, practically overnight.

If anyone else hit me with a claim like that, I’d be skeptical… but Larry’s a veteran trader and “Market Wizard” whose hedge fund was ranked multiple times in the top 1% by Barron’s. He’s well-known in the industry. There’s no one better for a strategy like this.

The Overnight AI Gains Summit will be next Wednesday, February 19, at 8 p.m. ET. You can click here to automatically add your name to the guest list if you’re interested in joining us…

It’s one of the biggest challenges in developing general-purpose, humanoid robots…

“Teaching” the robots to move with the kind of agility that comes second nature to us humans.

I’m sure we’ve all seen the videos of the jerky, awkward movements as robots struggle to do simple tasks. They tend to be funny, almost like slapstick comedy, because the failures aren’t from doing anything difficult.

But these types of failures are becoming far less common…

The combination of higher-performing actuators, more powerful semiconductors, and artificial intelligence (AI) – specifically neural networks and reinforcement learning – have completely changed the capabilities of humanoid robots…

The movements are – surprise, surprise – becoming a lot more human.

Cristiano Roboto

Earlier this month, a bunch of researchers from Carnegie Mellon University – with some collaboration with NVIDIA (NVDA) – released some cool research on this topic called ASAP – Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills.

The hardware (the robot) that was used in the research is from a private robotics company based in Hangzhou, China – Unitree.

Unitree was originally known for its quadruped robots. It released its first bipedal robot in 2023.  Its latest version, the G1, is shown below.

Unitree’s G1 humanoid robot | Source: Unitree

I’m noting this because the researchers basically used off-the-shelf humanoid robots that are priced starting price at $16K. These are university researchers with limited budgets, so they simply worked with existing robotic hardware – in this case, the Unitree G1.

Conceptually, the research approach was straightforward. The team ingested video of humans performing movement… and then that video was digitized in a simulated environment. Then, a form of AI – reinforcement learning – was used in the training environment to improve simulated performance.

Source:  Carnegie Mellon

After using reinforcement learning in simulation, that software output was applied to the Unitree G1 environment. And the final step was to deploy reinforcement learning on the actual real-world robot – to learn and improve the dexterity of movements.

This approach – of going from real-world video to learning/improving in a simulation to applying and learning/improving in the real world – proved to be highly effective in achieving agile motions that were previously difficult to achieve.

For example, a Cristiano Ronaldo move…

Or how about Kobe Bryant?

Or how about a 1.5-meter jump forward?

The research was impressive.

On a very limited budget, the researchers demonstrated more than a 50% reduction in motion tracking errors from simulation to real-world applications.

And clearly, from the videos above, an off-the-shelf robot was taught to perform some pretty complex movements.

What was just as interesting as the progress made was the limitations that the team identified in their research.

Fast Failure

As with any major tech development, fast failure is the name of the game. The researchers identified the following points of improvement:

• Hardware constraints – mainly that whole-body motions exert stress on the robots, which leads to motor overheating and mechanical failure (they broke two Unitree G1s during the research).
• To work properly, high-quality motion capture systems are necessary.
• Lots of video data is required for whole-body motions (i.e., just recording one or two videos of an action isn’t enough; hundreds are required).

To me, this explains why a company like Tesla (TSLA) designed its own custom hardware (actuators, motors, joints) – to overcome precisely the problems that the Carnegie Mellon team had with the Unitree G1.

And the high-quality motion capture systems and large video datasets for learning are simply a function of having the budget to collect that information. That’s an easy problem to solve for a corporation.

If a small team with a limited budget can make such impressive strides with reinforcement learning and off-the-shelf hardware, just imagine how quickly things are moving at Tesla with Optimus or Figure AI with Figure 02.

I’m expecting major updates from both companies on their humanoid robots – within the next two months.

The reality is that the feedback loop for training and learning complex human tasks on humanoid robots is becoming very short…

And that’s why the dawn of general-purpose humanoid robots is upon us.

What will you call yours?

Jeff Brownbot