Summary of Key Points
Tesla has converted the production lines for the discontinued Model S/X into mass-production lines for its Optimus humanoid robots. Optimus represents a crucial milestone in Tesla’s transition from a car company to an AI company, as it utilizes Tesla’s FSD (Full Self-Driving) software and some of its hardware. However, despite five years of development, mass production has yet to begin. While automakers have advantages in supply chain management and large-scale production, the “general capabilities” required for robots—such as picking up eggs or cleaning—differ significantly from those needed for autonomous driving, posing a major bottleneck in Tesla’s progress. Even experts like “Mutiujie” (a nickname for financial analysts) have admitted that they did not include the value of Optimus in their previous valuations of Tesla, and Elon Musk has criticized Wall Street for underestimating its potential. The reality is that developing general-purpose robots is more challenging than initially anticipated.
1. Optimus: Tesla’s Ace in AI Transformation—Why Is Mass Production Stuck?
Optimus is not just a casual project for Tesla; Musk describes it as a “true universal robot” capable of helping with everyday tasks. Experts recognize its potential, but why hasn’t mass production started after five years since the concept was unveiled? On one hand, Tesla shares factories and algorithms with cars, which seems to save costs. However, the needs of robots (being able to perform various tasks in different environments) differ greatly from those of cars. On the other hand, it’s often assumed that automakers understand robotics because of their expertise in hardware, but the core of general-purpose robots lies in software, an area where Tesla (and other automakers) have not yet made significant breakthroughs.
2. Algorithm Transfer: Copying Car Algorithms to Robots—A Close Call to a Mistake?
In its early stages, Tesla tried directly applying its FSD algorithm to Optimus. Former AI head Andrej Karpathy revealed that the robot initially thought it was a car because it used the same cameras, computers, and algorithms. The reason for this is that car algorithms are designed for safe movement (e.g., stopping at traffic lights, avoiding obstacles), while robot algorithms need to handle specific tasks (e.g., picking up a glass or handling an egg with precision). These algorithms serve different purposes, and simply reusing 60% of the FSD code does not address the fundamental issues robots face.
3. Automakers’ Hardware Advantages: Supply Chain and Production Expertise
Automakers have a clear advantage in hardware:
- Supply Chain Management: Many components (such as motors and reducers) are common to both cars and robots, allowing Tesla to leverage its large orders to negotiate lower prices and collaborate with suppliers to reduce costs. For example, Green Harmonic, a company that originally made robot reducers, was brought into Tesla’s supply chain.
- Large-Scale Production: Automakers have mature automated production lines and quality control systems, which improve efficiency and reduce costs, an advantage that startups (with manual production methods) cannot match.
However, these advantages only ensure that robots can be manufactured; they do not guarantee that the robots can perform their intended tasks effectively.
4. The Data Challenge: Teaching Robots to Work Like Humans
General-purpose robots need to learn practical skills, similar to how children learn through practice. However, gathering the necessary data is a significant challenge. Robots must understand things like gentle handling of objects (e.g., picking up an egg) or precise control (e.g., using a spatula). This problem is often referred to as the “Moravec Paradox.” To overcome this, robots need to be fed with extensive data on real-life tasks, such as thousands of attempts at picking up eggs while recording the necessary forces and angles. Automakers’ data from autonomous driving is not suitable for this purpose, as it relates to driving behavior rather than manual tasks. Currently, methods like using workers wearing cameras and sensors to record actions (e.g., cooking or packaging) are being used to collect data efficiently. Nevertheless, this process requires a long-term effort with no shortcuts.
Conclusion: Automakers’ Strengths and Weaknesses in Robotics
While automakers (such as Tesla and Xpeng) excel in hardware and production, they struggle with algorithms and data. General-purpose robots require advanced AI that can handle various tasks effectively, which relies on extensive real-life data and tailored algorithms—both of which are not readily available to them. Therefore, for Optimus to achieve mass production and practical usefulness, Tesla must overcome these weaknesses in software and data acquisition, a challenge that may be even more difficult than building cars.