Today, Tesla held their 2023 Annual Shareholder’s Meeting and while there was a lot discussed, one of the biggest themes was Teslabot. Today, Elon Musk confirmed that we will in fact get a Teslabot Podcast, something many of us are keen to subscribe to.
It is hoped that the podcast will take us inside the development challenges for the team and the difficult task of converting this prototype into a commercial reality over the next couple of years.
Teslabot timeline
2021: August – Elon Musk announces they will create a humanoid robot prototype due next year.
2022: October – Tesla shows Teslabot prototype walking at AI Day 2022 (V1 walking untethered, V2 standing tethered).
2023: March – Tesla showed off a number of robots, walking on their own, and importantly working together to achieve a task, along with detail around their work on custom-built actuators.
2023: May – During today’s annual shareholder meeting (around the 58min mark) Tesla showed off a new video of Teslabot, demonstrating their dramatic progress, in just a few months.
In the video we see 5 Teslabots, walking freely, and confidently and the big change for me was the refinement in cabling, now much of which has been miniaturised and integrated into the body.
As the bots walk, you notice just how much rotation is happening on the hip joints, this combined with smooth motion through their knee and ankle joints allows the bot to walk, while counter-balancing the upper torso.
If we look to the side of the body, we can easily see the outline of 4680 battery cells, used to power the electric robot.
As we look at the 5 robots together, not all are created equally. One of these features what I think is the next generation actuators in its thighs. These are finished in a silver metallic matching much of the rest of the body, compared to the other Teslabots that feature a black actuator in this position. This primary robot, let’s call him Sonny, also gets to wear a Tesla belt buckle, while the others are left with a Ken-style blank space that’d make Taylor Swift blush.
Tesla made a big deal about how much effort they are going to, to create their own hardware as off-the-shelf components, that could provide the characteristics they were looking for.
We see the Teslabots walk in an office environment (at least 3 out of the 5), then in a factory, where one day, they may be put to work to replace humans.
There’s also a segment of the 1-minute clip that shows the motor torque control, allowing an unattached leg to be sensitive enough to avoid crushing an egg.
In one scene, we see Teslabot walking outside for the first time (note: it’s a perfect day). This is labelled with ‘Environment discovery & memorization’.. which is followed by another shot from the computer vision system, clearly tracking the environment around Teslabot. Next, we see a really detailed point cloud of the environment Teslabot has navigated through.
If you ever needed a 3D map to be created and maintained, having a number of Teslabots walk through and scan an environment would be an incredibly efficient way to achieve that.
While Teslabot will feature the same Full Self Driving hardware chip as the vehicle, it may be that providing Optimus with memory is advantageous, while the car (for the most part) looks at the changing environment and adapts to it dynamically, every time.
While there are environmental variables and construction etc can occur, if you need to ask the Teslabot to go somewhere (even with geo-limits like the Tesla campus), then having a memory of where each building is located would be awfully handy for navigation. The key here is that once one Teslabot establishes this, other Teslabots can take advantage of it, assuming they are all constantly connected to Tesla’s data centre to leverage shared knowledge instantly.
Training Teslabot
Next, we finally get an answer to one of the biggest questions about Teslabot.. how are we going to train it?
Teslabot is effectively only as good as the skills it has to help us with boring, or dangerous jobs. To train Teslabot, we see Tesla leveraging human demonstrations and see a staff member in a motion capture suit.
Unlike most motion capture setups, there are no ping pong balls all over the suit, instead, the human is wearing a serious backpack, with a single tether (likely power), a headset with a complex camera array, likely mirroring the camera array inside Teslabot’s head. The trainer is also wearing a set of gloves and what appears to be a very precise tracing of each fingertip.
The trainer begins with a static pose, then bends forward to pick up each object with his left hand and place them in the black tray, then using his right hand, picks up each remaining object and places it in the silver tray. At the end of this training process, the trainer returns to his original static pose.
Later on in the clip, we actually see Teslabot then using that training to perform this exact task. Unfortunately, the single most important piece of information is left out.. how fast can Optimus aka Sonny learn? Were these clips shot just minutes apart, or days or even weeks apart?
We’re very used to new builds of software for cars taking a matter of weeks to rollout, but I expect and hope the development cycles for Teslabot are much more rapid, with so many more things on the to-do list, yet to learn.
What’s interesting to consider is if Tesla can commercialise the training setup and if they’re going to allow those who purchase Teslabot to train their own skills.
It is clear from watching this video (over and over at 0.25x speeds), the Teslabot prototype, already has an unbelievable range of motion and dexterity in the arms, wrists and fingers.. all critical and delivering on the objective, to replace human-designed tasks.
Special shout out to the 3 new Cybertrucks that are a nice backdrop to what I believe was the main event today, Teslabot.
Thankfully during Q&A, someone asked what I had hoped would be asked, that is where’s our Teslabot podcast?
Elon Musk confirmed that after each quarterly earnings call, the team will do a podcast, dedicated to the Teslabot for quote ‘anyone who cares’.
After having stated today that Teslabot could be responsible for more of Tesla’s future revenue than cars, I think plenty of people will be interested. Things got even more crazy when Elon suggested that each person on earth may want multiple Teslabots (I guess they’re free in this thought process), and that potentially as many as 10-20 Billion of them would be required.
Let’s start with selling your first Elon.
There is no doubt that a capable, easily trainable, affordable humanoid robot would be in high demand from businesses. With that demand likely to extend for many years, it’s not clear what the path forward is for consumer-level purchases and the price tag is definitely still unknown.
