This week, Elon Musk confirmed what I was already expecting, that at AI Day #2, we are going to see a working prototype. In my eyes, holding another AI Day and not talking about TeslaBot would have been incredibly strange, particularly given this will again be a pitch to the best engineers to come and work on the Tesla AI team, many of whom would be focused on the bot.
When the TeslaBot was first unveiled during the first AI Day (August 20th, 2021), it surprised the world, particularly given it wasn’t clear how this fit with the company’s mission.
Once we had time to think about the possibilities available to Tesla (and the world), through a smart, humanoid robot, it does make sense, particularly when Musk, and recently Drew Baglino, have spoken about the need to grow the business to an extreme scale over the next decade.
So if the bot really can be useful, first in Tesla’s factories, then to other businesses and even in our homes, what should we expect of the first working prototype, due on September 30th?
When the TeslaBot (codename: Optimus) was first announced, it was a stationary humanoid model, with a really refined, sleek exterior body. If we look at other humanoid robots, most are big, heavy, clunky implementations, very industrial, so seeing something more human-like, made the world sit up and pay attention.
This form factor is exciting, but only if Tesla can make the robot work, have intelligence and get more than 3 minutes of battery life.
When TeslaBot was revealed, they provided technical slides showing details about the bot’s internals. It will use 40 electromechanical actuators, 12 in the arms, 2 in the neck, 2 in the torso, 12 in the hands and 12 in the legs. This level of detail suggests they were a decent way along the development of the robot before it was announced.
Tesla also offered figures around its capabilities like being able to deadlift 150 pounds, move at up to 5MPH (likely software limited), and perform an arm extended lift of 10 pounds. These combined with other characteristics like being 5’8″ tall really felt like Tesla had done the work to ensure their electronics could fit inside the form factor.
While Tesla will borrow many parts from their vehicles for the TeslaBot, namely the HW3/HW4 computer, cameras and batteries, it’s not clear that they have the actuators to drop straight in. These could be developed in harmony with the Musk-owned SpaceX, along with lightweight, structural material for the skeleton and even plastic composites and fabrics for the exterior.
By the time September 30th rolls around this year, it’ll be close to 13 months since the announcement, so what should we expect at AI Day #2?
My initial guess was that Tesla was pretty far along when they announced TeslaBot, and I imagined that may have even involved a very early bot that was already walking. Given Elon’s comments, this week suggested they ‘may have an Optimus prototype’ that may have been far too ambitious for that time.
Right now, at the time of writing, we are 308 days since TeslaBot was announced to the world. I can imagine the engineers at Tesla have been working tirelessly to develop Optimus, but may have spent many of these past months on arms or legs independently. Once each of the appendages is working (fingers, wrist, elbow, shoulder), could it then be assembled to on the full-body, to work in harmony with the other body parts to attempt something like walking.
In terms of how Tesla would go about teaching the robot to walk, their game is obviously AI. Training humanoid robots on how to walk with AI is not new and something I believe Tesla could do quite quickly.
Here’s a great video from SIGGRAPH Asia 2019, showing a Deep Learning Framework known as Neural State Machine that allows a character to move naturally as it interacts with objects in the scene. It shows the character being able to pick up objects, move through gaps, around walls, as well as sophisticated moves like sitting down on a chair at a table.
Tesla’s AI team has access to massive computing resources and while much of that will be dedicated to their FSD efforts, if they’re serious about moving Bot from idea, to prototype to production in just a couple of years, they’ll need much more hardware, which is exactly why Dojo can help.
For me to be impressed by TeslaBot, I need to see it not just walk around in a fairly human-like way, but leverage its computer vision stack to accurately identify objects and do something useful with that information.
When we think about the factory applications for TeslaBot, it’s likely some of its first uses would be to move parts around the factory, ensuring employees working on the line always have a supply of materials to get cars completed as fast as possible.
If Tesla has the bot identifying and sorting objects, delivering different items to different locations based on what that object is, I’ll be really impressed. Throw a few curveballs to challenge its adaptability to ever-changing environments and I’ll be very impressed.
If it has the ability to learn a new skill on the fly, I’ll be absolutely stunned, bonus points if the voice recognition is working. Musk has explained a core safety principle would see the Bot respond to anyone say ‘stop, stop, stop’, which begs the question, how do you avoid anyone yelling that during the demo?
The big hope that I have, is that Tesla shows multiple prototypes on stage in September, and they are able to work together to accomplish a task faster than a human. Then we need to talk about the cost, battery life and charging story.
The battery life and time to recharge are critical to the economics of this being better than a human (or shift of humans). So far at least, it seems the 4680 battery is the most likely, given the energy density requirements are so high (lots of power in a small form factor).
The ramp of the 4680 in Texas appears to be slower than expected, so if the TeslaBot does rely on them, that’s something Tesla will need to explain if we’re going to see thousands of bots compete for precious batteries that could be going to cars or stationary storage.
There’s a lot to look forward to with the TeslaBot, but if it falters and was to fall over, or mess up in a major way, the world would rapidly reset its expectation on when these will actually make it to market.