Tesla’s Optimus Hits Snags in Hands, Suppliers as Scale-Up Begins

Tesla has ramped up production of its Optimus humanoid robot roughly tenfold in recent months, but the company is still struggling to manufacture the machines reliably at scale. Its production lines are grappling with problems involving the robot’s intricate hands, automated equipment and supplier constraints, according to people familiar with the project.

Last month, Tesla produced several hundred robots a week, up from a few dozen robots per week during the second quarter when the company was doing small batch testing for production, according to two people familiar with the work. Managers have told staffers the company wants to set up a continuous automated production line capable of producing more than a thousand robots a week by the end of the year, the people said. That is still far short of its eventual plan to produce around 20,000 a week.

Tesla CEO Elon Musk has repeatedly characterized Optimus as the company’s most important product—pitching the humanoid robot as likely to take over everything from manufacturing and data center upkeep to everyday tasks like cooking and cleaning. Tesla primarily earns its revenue through vehicle sales, but Musk has said Optimus will help Tesla pivot further into becoming an AI company.

Tesla has repeatedly shifted the timeline for the robot. In 2022, less than a year after Musk first announced plans to build a humanoid robot, he said the robot could be production-ready as soon as the following year. In January, Musk said during the World Economic Forum that Tesla could begin selling Optimus by late 2027. In April, he said “we should be able to show” off the upcoming version of Optimus by the middle of the year,” but that date passed with no demonstration.

To be sure, Musk has also been clear that manufacturing Optimus “is a very complex problem to solve,” as he said in July on an earnings call, adding that “the production scaling challenge is very substantial.”

At the time, he said that Optimus production would start “soon.”

Much still has to be done to produce a robot Tesla can sell. The robots coming off its Fremont, Calif., factory line—known as Optimus V3—are not the exact version the company plans to sell to customers, the people said. The company is still fine tuning that version, which will have to meet stricter durability and reliability thresholds before Tesla offers it commercially.

The V3 is lighter, equipped with more cameras and designed to look more human than earlier versions of the robot, two people familiar with the design said.

Most of the robots now being made are being used internally for testing, training and data collection, including at Tesla’s offices in Palo Alto, Calif. At its factories, Tesla uses the robots in tightly controlled areas where their movements and tasks can be closely supervised, according to one of the people. In these settings, the robots are programmed for specific tasks rather than operating as generalized or remotely controlled machines, the person said.

The company plans to gradually expand the robots’ operating areas as their performance becomes more reliable over time, they said.

Tesla plans to lease the robots initially to outside commercial customers rather than sell them, allowing the company to retrieve, upgrade or refurbish the machines after the lease period, one person said. The company also plans to use data collected by the robot in customers’ factories and warehouses to improve its AI system, they said.

Tesla has drawn up a short list of companies whose operations might make them suitable customers for the robots, typically businesses that have factory or warehouse setups similar to Tesla’s facilities, which would make it easier for the robots to adapt to their new environments.

Tesla did not respond to a request for comment.

Cars to Robots

In early May, Tesla shut down the Fremont production lines for its Model S and Model X vehicles after the company discontinued the vehicles and began moving some of its factory workers and engineers onto the test line for Optimus, another person familiar with the work said. It also moved dozens of engineers from its Model Y line to help, they said.

But as production ramped up, problems began to emerge. At several stations on the production line, including the ones that assemble the robot’s hands and joints and those that install software and test its electronics, the equipment that holds parts in place can’t line them up precisely every time. Not only are the Optimus production lines new, but the Optimus parts are much smaller and have to fit together far more precisely than car parts.

The end result is that more robots coming off the line need fixing, the people said. Moreover, some newly developed tooling and automated equipment doesn’t always work properly when the line runs fast, one of the people said.

One of the most difficult aspects of production has been building Optimus’ hands so they can be manufactured at scale. Musk has said the hands will have the same dexterity as a human’s. The hand and forearm have more than 100 screws and other small components that currently require a worker to assemble manually, one of the people said.

It’s also proving difficult to make the hand durable enough for sustained use. Some of its touch sensors have had reliability problems, one of the people said. Tesla has developed a “sensing glove” containing the hand’s touch sensors that can be replaced without requiring replacement of the entire hand, and it plans to incorporate that system into the model next year, the person said.

The complexity of the hand design has also contributed to delays and higher costs for the newest version of the robot, according to people familiar with the program. Musk has said Tesla plans to eventually sell Optimus to the public at a price point between $20,000 and $30,000.

Tesla is continuing to tweak the design in parallel with construction of the production lines to make sure the product can be built at scale, two people said.

Musk has acknowledged the challenges. In July, on an earnings call, he said Optimus will be the hardest Tesla product to scale because the company has had to build much of its supply chain from scratch. Commercial humanoid robotics is a developing industry that does not yet have an established network of suppliers.

While Tesla is developing much of the technology in-house, the company is relying on outside suppliers for some components, primarily motors and the precision gears that move the robot’s joints.

Much of that work is done in China. Many of the parts that go into humanoid robots are made there, including the motors and sensors, the parts that move the robots’ joints. Some U.S. robotics startups are so dependent on China that they have enlisted people to carry parts from the country to the U.S. in their luggage, The Information reported.

Tesla is testing and verifying new suppliers while pushing some to build factories outside China, and it’s offering bigger orders to those that do, the person said.

But while some suppliers can produce components successfully in prototypes or small batches, they have struggled to maintain quality and consistency at higher volumes, according to a person familiar with the program.

Basic Tasks

Another hurdle: The robot’s AI is also not yet capable of reliably handling a wide range of tasks, according to three people familiar with the system. For now, the robot relies heavily on being trained and programmed for specific tasks, the people said. When faced with a task or situation it has not been trained for, its performance can be unpredictable.

Tesla is trying to make Optimus adaptable to new environments by teaching it a library of basic movements that can be combined for different tasks, rather than training it from scratch for every job, according to multiple people familiar with the training. But the robots can still take several days to learn even basic tasks.

Training a robot’s AI requires an enormous amount of data. For several years, Tesla’s team of data collection operators has been collecting data through remotely operating the robot. Workers wear suits designed to track minute movements, or they record video of colleagues performing tasks such as folding a T-shirt, wiping a table or sorting batteries, to teach Optimus how to perform those tasks.

Tesla currently has more than 500,000 hours of training data and aims to double that amount by the end of the year, one person familiar with the efforts said.

With that in mind, Tesla has steadily expanded the number of personnel devoted to data collection. In January, the company told workers at its factories in Texas and California it wanted some of them to wear data collection suits that would record their movements to teach the robot how to perform tasks on the factory floor, according to a recording of the meeting.

But production workers knew the robots were designed to eventually replace them. Some complained. Tesla began hiring dedicated data collectors to perform tasks on the floor and setting up cameras around the factory to collect data, three people familiar with the work said.

It has conducted much of the data collection work in areas separate from active production lines, two of the people said. The workers perform tasks such as sorting parts while the cameras on their helmets and other cameras stationed in the factory record their movements, the people said.

In the spring, meanwhile, a large portion of Tesla’s data annotation team working on self-driving software shifted to Optimus work, one person said.

Then, over the summer, Tesla began establishing training hubs across the country, including in Colorado, Arizona and Florida. The teams consist of dozens of data collectors who use helmets equipped with cameras and heavy backpacks connected to computers to capture their movements and generate first-person visual data, sometimes with haptic gloves, which track hand movements, five people with knowledge of the work said.

The company’s goal is to eventually create a feedback loop in which robots deployed in the real world generate data Tesla can use to improve the software. It can then push those improvements back to the fleet, similar to how the company has trained its driver assist software using millions of customer vehicles, three people said.