The Future Of Tesla is Robotics

Image licensed to Phil Whelan from Besjunior on Envato Elements

About six years ago, I was having my regular six weekly chat with my hairdresser about Tesla. We were both enthusiastic about the company and both held stock. I was trying to convince him that Tesla wasn’t a car company. At its core, I told him, I believed Tesla was actually a battery company, since batteries were key to everything they did and where they needed to focus on if they wanted to scale.

Batteries have since proven to be key to Tesla’s success and the biggest bottleneck in scaling the company. At Tesla’s “Battery Day” and other company events, Tesla has clarified that they understand that scaling battery production is key to their success and that they are taking steps to ensure their own destiny. This includes addressing everything in the makeup of a battery, right down to the mining of the raw materials, such as Lithium. There is no doubt that Tesla is, and probably always will be, a battery company as you project into the future. But I think robotics will start to overshadow that in the next decade or so.

Software 2.0

Tesla’s mission to accelerate the world’s transition to sustainable energy drives their battery ambitions. Driving down the cost and driving up the scale of batteries is not easy. It looks likely that they will dominate this space and hence have an advantage to continue to dominate the EV market, but the margins they get from batteries will never be exciting. Tesla leverages its battery efficiency to sell electric vehicles and leverages its electric vehicles to sell software. Software is where the real margins are. Tesla is good at software, but has not yet got into its stride in terms of selling software. But that’s coming.

As Director of AI at Tesla, Andrej Karpathy, has said for many years that he believes AI, or Machine Learning, is “Software 2.0”. In their work on autonomous driving, they’ve seen Machine Learning slowly consume more and more of their traditional software. This means instead of painstaking writing each line of code and manually figuring out all the edges of a problem, you simply get an exhaustive set of examples and have the machine deduce the rules itself. Tesla’s work on autonomous vehicles is one of the world’s hardest Machine Learning problems to date and, as it scales with data, engineering and infrastructure, it’s positioning Tesla to be a leader in Software 2.0.

Andrej Karpathy’s Twitter comments on the impressive progress of NeRF (Neural Radiance Fields)

Flywheel

There are over a million Tesla vehicles on the road. Each of these is a node in a vast Machine Learning architecture, receiving requests for video clips. They continuously scan their environment for examples of obscure, but specific, edge-cases to report back to the mothership. The scale of this means a rapid feedback loop on development of any edge case that has been manually reported, detected via human intervention, or by the software seeing something it didn’t expect (e.g. the human driver continued when the observing computer detected they should have stopped).

Self-driving Machine Learning models are being continuously built and deployed to this rapidly growing fleet. These models run within the car on incredibly fast custom hardware. They consume data from cameras around the car to track all objects, predicting what they are and where they might move next. They understand how different objects behave, what they mean, or how they interact with each other. For instance, a complex sets of traffic lights and accompanying road markings, or the position and wording on signs (“STOP! Unless turning right”). They track objects that are temporarily out of view. The models are even capable of looking at pixels from cameras and determining how far away they are.

These models take in the world through inexpensive cameras, understand it, predict the future, plan, and execute their next moves using Tesla’s motors and those highly sought-over batteries. A Tesla vehicle is simply a robot that has an appetite for humans, but like most robots is unable to digest them.

Cars are cheap. Any fool can make a car. Reliable, safe and fast cars are harder. They cost money. But a robot car that can drive itself, pick you up and drop you off, or go and collect something for you? A car that can drive around as an autonomous taxi making hundreds of dollars a day? That’s going to be pretty expensive. But not to Tesla. Remember those inexpensive cameras, that in-house custom hardware, that software they built themselves? That sounds like significant margin to me.

From listening to Tesla over the years, it sounds like their battery split is generally targeted for 50:50 between electric vehicles and stationary energy storage systems. So far, that seems to have leaned more towards vehicles. The availability of batteries has been a limiting factor year after years. From a purely profit point-of-view it seems hard to justify the small margin on stationary storage compared to that of the high margin from selling their robotic vehicles. But a sustainable future requires stationary storage and as long as the mission doesn’t change, rolling out more stationary storage will be a priority.

It’s also worth noting that there are some monetization strategies through software for stationary storage in the form of AutoBidder.

Robots In Manufacturing

Hopefully you agree with me that Tesla is building robots. They just happen to look like, and provide the utility of, consumer automobiles. But just as an iPhone could be mistaken for a Nokia (if you squint from a distance), the utility of a Tesla vehicle will evolve and open up new and unimagined utility in the future.

Like Apple, Tesla controls both the software and the hardware. This provides incredible opportunity for not only cost savings, but great user experience, innovation and rapid evolution.

Unlike Apple, Tesla is also manufacturing their own products and, as mentioned, will soon be getting into mining for materials. Robotics is key to modern manufacturing.

Unfortunately, when Tesla was ramping Model 3 production and seeing robots as the future of manufacturing, they jumped in with both feet, only to find the water was too shallow. They needed to quickly scale back their excessive robotic usage in their production line. They also found themselves rewriting some of the software that controlled the robots they had purchased. They’d hit the limitations of what industrial robots could offer.

The Next Robots

On the cusp of conquering autonomous driving, surely Tesla must be looking at what their next robotic venture will be. They have the vision systems, motors and batteries. They have a growing competence in manufacturing. So what’s left?

It’s true that the vision systems of electric vehicles will not have had experience with objects encountered beyond the streets, such as stairs or teapots (sorry, I’m British). But Tesla’s internal progress in AI is unlikely to be measured by the linear rate in which they can recognize new objects. More likely they measure the speed at which an edge case is detected, a volume of applicable high quality data is retrieved, and a new model gets into production (or test vehicles). This, plus constant evolution of algorithms, means exponential increases by any other measure.

Algorithms needed for controlling robots used in manufacturing will be different from that of electric vehicles, but there’s no barriers to Tesla acquiring this knowledge. I’m sure any expert in their field would be more than willing to consider seeing their expertise productionized by Tesla.

Conclusion

Tesla’s focus right now is on ramping up the number of electric vehicles it puts onto the streets. That’s the number most people care about. Meanwhile, there is rapid progress on the autonomous side of Tesla. Combined, this can be seen as a single robotics product. The amount of investment and energy going into Tesla’s AI should not be underestimated.

Once the product of the autonomous vehicle is “complete”, the growing engineering prowess behind it needs to be applied elsewhere. Tesla needs more automation in its factories. They need more automation in their mining efforts. So it seems natural these would be their next stops. But why stop there? As they build out an autonomous robotics platform within Tesla, the cost for productionizing new high-margin robots drops.

Tesla might always be “a battery company”, but robots are their future.

Software Engineer