Managing Editor’s Note: As longtime readers know, Jeff’s been two steps ahead of the rise of autonomous vehicles for years.
And now, just a couple months away from Tesla’s robotaxi rollout in June, he’s hosting an event to help his readers get ahead of the curve.
He’s getting into all the details – including a handful of small companies you should be paying attention to – next Thursday, April 10, at 1 p.m. ET. You can go here to automatically sign up to attend.
One of the major challenges in building general-purpose humanoid robots right now is mastering humanlike motion.
It’s a complex problem to solve, as it involves both hardware – the actuators that are mechanical devices that turn energy into motion – and software, the artificial intelligence (AI) to precisely control the actuators to achieve its task.
Humans are remarkably efficient when it comes to walking, running, or manipulating objects. We can perform these tasks quickly, and they feel like second nature, which makes sense… as we have practiced these tasks since childhood and reinforced how to do these tasks over years if not decades (not accounting for the fine-tuning of evolution, obviously).
A humanoid robot’s gait or motion is slower. While capable of performing tasks with precision around the clock (other than charging time), human productivity has been superior to humanoid robots.
Up until now.
Below is a short clip of the kind of slow, mechanical gait that has been typical of humanoid robots.
Agility Robotics’ Digit Circa August 2024 | Source: Agility Robotics
But this kind of movement has quickly become a thing of the past.
Advancements in both actuator technology and AI are solving this challenge, which will flip the productivity advantage to the robots.
The last few days have seen an odd confluence of announcements from a few of the major humanoid robotics companies. They have all made major, and visible, strides, pun intended.
Figure AI highlighted its latest improvements, which it calls “natural humanoid walk” for humanoid locomotion. The robotics company has improved its end-to-end neural network by using reinforcement learning to better master the task of locomotion.
Below is a useful video comparing the “old” Figure 02 gait with the new and improved locomotion that leveraged reinforcement learning for a more human-like walk.
Figure 02’s Humanlike Walk | Source: Figure AI
Figure AI created a high-fidelity, physics-based simulation where it could simulate years of training in just a few hours. Shown below is a short video representing a simulation running thousands of virtual humanoid robots under a variety of parameters and scenarios.
Figure AI Physics-Based Humanoid Simulation | Source: Figure AI
In this kind of simulation, virtual humanoid robots can be trained on all types of terrains, slopes, and surfaces. And they can be “tested” on environmental events like slips, trips, and shoves.
And then, as if by magic, real-world humanoid robots can simply “download” this newfound collective understanding – all gained from simulation.
Without today’s latest generation of GPU technology, this kind of high-fidelity, physics-based simulation wouldn’t be possible. Today, computational power and simulation software have improved so much that transferring the learned knowledge from simulation to the real world is proving to be highly effective.
A Team of Figure 02 Running on Same Neural Network | Source: Figure AI
Shown above is a number of Figure 02 robots all running on the same neural network from training that took place in a simulation. The AI running these robots will, of course, benefit from additional real-world data, which will be fed into the reinforcement learning for even better human-like performance.
Seeing images like this makes me wonder, what will we call “them” when they are in numbers?
I have no idea, but we’re going to have to call them something. They’re coming, and fast.
I suspect what we call them will depend on the application. For military – squad, squadron, legion, army, or swarm might make sense. For agricultural applications – a farm or herd would sound logical. For enterprise applications – a fleet or pod of robots would make sense.
What’s more important for us to understand is that these technological approaches are well-known and available to the rest of the industry. It’s clear that reinforcement learning has proven to be very effective, so it’s no surprise to see similar advancements from other major players.
Below is a very recent video from Boston Dynamics and its latest generation of its Atlas robot.
Boston Dynamics’ Atlas | Source: Boston Dynamics
Notice the natural gait in Atlas’ walk, and even more impressive is how Atlas is now running.
This is a major improvement by Boston Dynamics from just six months ago, almost hard to believe.
And check out this clip of Tesla’s Optimus just released yesterday. Notice anything different?
Tesla Optimus’ New Strut | Source: Tesla
Notice Optimus’ arms, how they swing and bend naturally at the elbow. Very humanlike, and completely different from the stiff-armed Figure 02, Digit, and Atlas humanoid robots.
October was the last major update from Tesla on Optimus at its We, Robot event. It was just a glimpse of what Tesla was working on in the background. I expect a larger update on Tesla’s Gen 3 Optimus within a month or two…
And I suspect, as usual, Optimus is a step or two ahead.
Jeff
Optimus in the Office | Source: Tesla
The Bleeding Edge is the only free newsletter that delivers daily insights and information from the high-tech world as well as topics and trends relevant to investments.
The Bleeding Edge is the only free newsletter that delivers daily insights and information from the high-tech world as well as topics and trends relevant to investments.
