On Tesla’s 2022 AI Day, CEO Elon Musk presented an update on the company’s Optimus robot. And like many product reveals that Musk makes, the humanoid robot was met with a wide variety of reactions, ranging from awe to ridicule.
Tesla employs some of the brightest engineers and scientists, and Musk has a track record of accomplishing things that seemed impossible. So, I wouldn’t be quick to dismiss what Musk and his engineers say on stage.
That said, it is really hard to evaluate the robot from a limited on-stage demo or the pre-recorded videos that were shown during the presentation. And like many other companies that dabble in artificial intelligence and robotics, Tesla has a history of overpromising and underdelivering.
Here are some of my own thoughts on Optimus, and some interesting facts that weren’t covered by the media.
Not cutting edge, but impressive nonetheless
Many people have pointed out that Tesla’s Optimus is not nearly as advanced or impressive as other humanoid robots. Clearly, at this stage, Boston Dynamics Atlas is far more capable than Optimus, based on what both companies have presented so far. And there are other companies and labs that have created two-legged robots that are better than Optimus.
The best technical discussion I found on the topic was this Twitter thread by robotics professor Christian Hubicki, in which he provided a very fair assessment of Optimus, based on what we’ve seen so far.
“Am I blown away? No. Am I laughing? No,” Hubicki wrote. “First, the team did a good job. They came a long way in about a year(?), going from zero-to-robot from the ground up. Also, doing a live demo without a tether (safety catch rope) is braver than people know.”
Hubicki that the reliability of the robot remains to be seen.
Most other companies that Tesla is being compared to have been in the industry for a decade or more. Tesla has managed to reach this point in a little over a year.
What is more interesting is that, if their estimates are correct, they will be able to produce their robots at under $20,000, which will be an impressive feat even if the retail price is twice that amount. For comparison, the four-legged Boston Dynamics Spot, which is much less sophisticated than a humanoid, is sold at $74,000. And according to some estimates, robots like Atlas costs more than a million dollars to produce.
Cool reuse of existing technology
One of the things that impressed me was Tesla’s reuse of its existing hardware and software to create the Optimus robot. Optimus is powered by the same neural networks used in Tesla’s Autopilot self-driving technology and the System-on-Chip (SoC) in Tesla cars. Tesla has long been a detractor of lidars and uses pure computer vision in its self-driving technology, so I suppose they will continue to use the same approach in Optimus. And using a processor that has been designed for battery efficiency will help improve the battery life of the robot. According to Tesla’s presentation, Optimus can work for an entire day with a single charge of its 2.7kwh battery pack (again, Tesla has been able to make use of its unique experience in developing batteries).
The neural networks obviously need repurposing and retraining before being transferred from self-driving cars to humanoid robots. But the modular approach Tesla used in its deep learning architecture has probably made it easier to retrain parts of the network instead of doing full end-to-end retraining.
One thing I’m curious to know is how they obtain the data needed to fine-tune their deep learning model. Tesla’s self-driving technology relies heavily on the tons of data that the company gathers through the cars it has sold. Humanoid robots will be working in fundamentally different environments and will need different training data. Advances in simulation engines are making it easier for research labs to train robots with minimal real-world data. But clearly, Tesla (still) doesn’t have the advantage here.
In the demo, the scientists said that the neural network was retrained on “the new platform” and also mentioned neural radiance fields (NeRF), a deep learning technique for creating 3D scenes from 2D images. But there is no further detail on how much simulation versus real data they are using.
And Tesla is opting for a full human-like body with five fingers and opposing thumbs. The hands will have 11 degrees of freedom, which is short of the 27 DoF of human hands but a very challenging feat nonetheless and the reason why many robotics companies are using more simplified architectures. This means that training the robotic hands will be much more challenging and will require a lot more data. They have used motion capture to create a repertoire of different poses and moves, which the robot then adapts to the task it wants to perform (this is similar to what the Atlas team is doing).
The current demo shows that the robot’s hands are still shaky, and I’d doubt that it can handle objects that require dexterity and a ginger touch. I’m interested to know how they plan to bridge that gap.
Level of autonomy and cognition
Another question that remains is the level of autonomy that Optimus will have. Will it be a teleoperated robot, or will it be a fully autonomous system that can be given broadly defined tasks and expected to carry them out without a remote human operator?
The Optimus demo shows the robot picking up objects and carrying out tasks such as watering plants. In the demo, the object detection and image segmentation seem to be working flawlessly. But detecting objects in an environment is just a small part of the challenge of robotics. The robot must also be able to map its environment, plan its route, prioritize its objectives, deal with sudden obstacles and interruptions, and perform a lot of more challenging tasks to be perfectly autonomous. These are areas that still don’t have perfect solutions.
Tesla has managed to make impressive advances in some of these areas in its self-driving technology, which can be very helpful in the development of Optimus. However, the environments of humanoid robots are much more unpredictable and difficult to master. Unlike self-driving cars, which are trained on avoiding collisions with humans, humanoid robots are expected to work with humans, which will require much higher accuracy and safety standards. And their environments are much more diverse and unpredictable.
Personally, I’m a fan of Spot’s semi-automated model, where a human operator specifies a trajectory or waypoint or series of tasks, and the robot navigates the path while detecting and avoiding obstacles. It provides the best combination of advances in AI/robotics and human intelligence. It will be interesting to see what kind of labor distribution Tesla plans with its robots.
Why would you need a humanoid robot?
In his presentation, Musk suggested that millions of Optimus robots will eventually be produced. Given Tesla’s experience in sophisticated manufacturing and production lines, I wouldn’t be surprised if the company manages to reach such output levels and reduce the costs of producing humanoid robots through economies of scale.
But I’m still puzzled at what will the use cases of humanoid robots will be. Musk was not very specific on the uses of Optimus aside from its potential to automate manual labor.
Musk is not new to robotics. Tesla uses robots to automate the manufacturing of its cars, and it is doing it very efficiently. And in manufacturing, warehousing, construction, and areas where physical labor is required, the better solution is to create robots for the specific tasks and to shape the environment to put the robot to better use.
That said, mobile robots have grown more capable in recent years, opening the way for more dynamic applications and use cases. For example, Boston Dynamics Stretch is suitable for deployment in flat-floored warehouses where the arrangement of objects constantly changes. Spot, on the other hand, is a four-legged robot that can navigate rough terrain and work in environments that are dangerous for humans, such as mines and industrial complexes.
Therefore, I don’t foresee the future as a bunch of humanoid robots replacing humans in a lot of tasks. Instead, I expect robots to appear in various forms (on wheels, legs, drones, etc.) taking on specific tasks, cooperating with humans, and gradually becoming capable enough to do most of the work.
However, history has shown that we often fail to predict the future use of new technologies. Just look at the history of tech luminaries underestimating the internet, mobile computing, smartphones, social media, and other technologies that have become inextricable parts of our lives. Who knows what humanoid robots are destined to accomplish?