Is Elon Musk's Optimus robot on a stable footing?

Elon Musk envisions Optimus as a robot that can walk steadily and use its dexterous hands to perform various complex and non - standard tasks, entering factory assembly lines or home kitchens. However, to achieve this goal, there are several major challenges, including meeting the standards for balanced motion control, autonomously completing generalized tasks, and achieving reliability close to automotive - grade.

But in reality, in January 2024, a video of Optimus folding clothes was released. In October of the same year, at the "ME Robot" press conference, Optimus walked into the crowd and interacted with them, but it was actually remotely controlled. Another example is the opening of the Tesla Diner, where we saw a remotely controlled Optimus in charge of the "popcorn machine". These have been "criticized" in the industry as the results of remote - control operations.

Even when compared with domestic robot companies, in terms of motion control capabilities, domestic second - tier companies like Yuejiang and Xingdong Jiyuan may have a slight edge. Currently, Optimus does not show obvious advantages either in lower - limb walking ability or upper - limb grasping ability.

Does this mean that Tesla's robots are no good? Of course not. Tesla's advantage lies in full - scale self - research and software - hardware synergy. For example, at the actuator level, Optimus has integrated all 28 joints into "integrated electric drive modules", and has self - developed motors, reducers, encoders, drive boards, and thermal management systems, which is a typical "vertical integration" approach. This is significantly different from the practice of some startup robot companies that purchase motors/reducers externally and only focus on control integration.

Image source: youtube

In terms of batteries, Tesla uses its own 4680 cylindrical cells and a self - developed BMS. Startup robot companies usually buy third - party power lithium batteries. Especially in the field of chips, most other robot companies directly purchase chips from NVIDIA or Intel, while Tesla has independently developed the FSD AI5 SOC chip, which is specifically adapted to the pure - vision end - to - end model.

Image source: AI

Interestingly, domestic robot companies often benchmark against the Optimus robot. A media once interviewed an insider at Zhiyuan Robotics and asked for their opinion on Tesla's robots. The insider said that if Tesla's robot could score 7 points (out of 10), then Zhiyuan might only score 3 - 4 points at this stage, indicating a very obvious gap. It should be noted that this was the judgment before the third - generation update of Musk's robot.

The biggest update to the third - generation Optimus is actually the exponential growth of its generalization ability. Tesla is shifting from data collection methods using "motion capture suits and VR headsets/remote control" to more "human videos", and directly using human videos to train the robot's model.

Image source: tesla

In contrast, once a general visual system is successfully trained, in principle, it can be deployed anywhere on Earth because it does not rely on any pre - existing environmental maps. In terms of cost, the cost of a multi - sensor suite can be as high as tens of thousands of dollars, while the cost of a camera is only a few hundred dollars. This huge cost difference gives the pure - vision solution a natural advantage in commercialization and large - scale deployment, enabling advanced driving assistance functions to be promoted to every consumer - grade vehicle.

However, this is an ideal state. If everything were really that smooth, given Musk's personality, Tesla would surely mass - produce 10,000 units this year and start with 500,000 units next year. But obviously, Musk has not done so because there are still many fundamental problems to be solved.

According to the latest research, 2D videos alone are not sufficient to meet the training needs of embodied models. To make embodied models have better generalization and higher task success rates, multi - modal data such as the 3D coordinates and trajectories of the robot's end - effector, 6DOF poses, the trajectories of the manipulated objects, and first - person view videos are also required. Currently, this type of data still needs to be collected through expensive manual tele - operation methods and cannot be replaced by 2D videos.

In addition, current dexterous hands have difficulty dealing with complex and changeable physical scenarios (such as grasping tasks in different shapes and environments), including Tesla's current dexterous hands. The fundamental problem lies in the dual lack of data volume and coverage. Moreover, high - quality data is more crucial than large - scale data, which directly determines the model's efficiency and generalization ability. In this regard, Tesla is not far ahead of its competitors.

Is it a bubble or a once - in - a - century opportunity?

Tesla's critics believe that Tesla is a company that once created miracles and significantly accelerated the world's transition to electric transportation. Then, its CEO went crazy, sales began to decline, and profits started to drop. To maintain the absurd stock price, the CEO decided to ride on the AI bubble, and that's all.

Fred Lambert, a commentator for the US media Electrek, said that the plan "is just a hodgepodge of vague AI promises" and even mocked its humanoid robot for "not even being able to hold popcorn properly". Compared with Musk's previous clear and executable master plans, the fourth chapter is more like "opium for Tesla shareholders", aiming to make them indulge in the fantasy of "infinite growth" of AI stocks without actual support.

However, this is not the first time Musk has made wild claims. The completion rate of the master plan in 2016 was also not very high.

The master plan in 2016:

1. Create amazing solar roofs that seamlessly integrate with energy - storage batteries.

2. Expand the electric vehicle product line to cover all major market segments.

3. Develop autonomous driving capabilities that are 10 times safer than human manual driving through large - scale fleet learning.

4. Let your car earn money for you when you're not using it.

Although the solar roofs and batteries in the first item have started to make a profit, they are still far from the demonstrated effect. In the second item, Tesla has indeed expanded its electric vehicle product line to most consumer segments suitable for the mass market, but the Model 2 for the low - cost market has never been launched. The third and fourth items can be regarded as the same thing, which is autonomous driving. The result is that Musk's robotaxi is still in the testing phase, and full - scale autonomous driving has not been achieved. As a result, this goal has been carried over to this year's master plan.

Image source: tesla

It is written in "Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future" that once Musk enters the "full - throttle" mode, it means that he will focus on a certain business for a period of time. This business could have been promoting mass - production of cars or redesigning and producing the Raptor engine before, and now it might be Optimus (which is the wish of all fans). After all, 80% of Tesla's value comes from Optimus.

Image source: the Internet

Should we believe Musk's stories? Musk often says, "If your dreams don't scare you, you're not dreaming big enough." Although reusable rockets existed before Musk, the cost was still very high. It was not until Musk's Falcon 9 rocket that the cost of space launches was significantly reduced, and its single - launch cost has become the lowest in the world. Tesla's cars were on the verge of collapse several times. It was Musk who slept in a factory tent and redesigned the production line from scratch. This was not only a revolution in the automotive industry but also a major change in production methods and efficiency, which led to the increase in annual sales from a few hundred to nearly two million units.

Musk almost never plays by the rules. Recently, Tesla's board of directors offered Musk a compensation package worth $1 trillion, requiring him to increase the company's market value from $1 trillion to $8.5 trillion in 10 years, deliver 20 million cars, scale the robotaxi business to 1 million vehicles, and deliver 1 million Optimus robots.

Image source: Bloomberg

This means that Tesla needs an average annual compound growth rate of 26%, and its market value would reach the equivalent of two NVIDIA plus half of Microsoft, which is unprecedented. Musk has welcomed this, although the compensation package still needs to be voted on by the board of directors in November.

Of course, just making all kinds of "promises" will not convince the board of directors and shareholders. So, Musk recently spent $1 billion to buy 2.57 million shares of Tesla stock. The market responded quickly and positively. After the disclosure of the document, Tesla's stock price soared by more than 8% in pre - market trading on Monday, reaching the $428 level, boosting the company's market value by about $100 billion.

At critical moments, it still has to be Musk. Robyn Denholm, the chairman of Tesla, said in a media interview that Musk is a "once - in - a - generation leader" and emphasized that retaining him is "the best way to optimize Tesla's future".

Although the talk of an AI bubble has been around for several years, we haven't seen the bubble burst. On the contrary, leading players are increasing their investment in AI, and some performance has started to be realized. The most fatal bubble is never excessive leverage but unfulfillable performance, which is just a game of passing the buck.

Whether it's AI or robotics, it's still too early to talk about the end - game. What we should do now is to have more people and capital pour in, build a solid infrastructure, make application preparations in various industries, and dare to pursue the dream of a trillion - dollar market.

References

Elon Musk shows his commitment with real money. Source: Wall Street Journal

Did Elon Musk announce the abandonment of remote control? Source: Robot Outlook

Who is the real leader in the humanoid robot field? Source: Silicon Steps Run

Is it true that Tesla has abandoned remote operation? Source: Embodied Era

This article is from the WeChat official account "Chief Business Review" (ID: CHReview). Author: Zuo Jing Guantian. Republished by 36Kr with permission.