Tesla is teaching its Optimus humanoid robots partly by recording how humans do physical work. Ars Technica reports that some workers asked to provide those demonstrations are reluctant, while the company targets 1,000 robots a week by the end of 2026.
Key takeaways
- Ars Technica reports that some Tesla workers have resisted taking part in training the company’s Optimus humanoid robots, on the view that they are teaching a system intended to do their jobs.
- Modern humanoid robots are largely trained by imitation, which means the data they need has to be produced by people performing tasks in the physical world.
- Ars Technica reports that Tesla is aiming to build 1,000 Optimus units per week by the end of 2026, a rate that would imply roughly 50,000 a year if sustained.
- The number of Optimus units built so far, the scale of the reported worker reluctance and Tesla’s response to it are not established by the available material.
Optimus, Tesla’s humanoid robot programme
Optimus is Tesla’s general-purpose humanoid robot: a bipedal machine with two arms and articulated hands, intended to perform physical tasks in environments built for people rather than for fixed industrial arms. Tesla has shown prototypes publicly at company events since the early 2020s, and the programme has been positioned as an extension of the company’s work on driver-assistance software, reusing the same broad approach of training neural networks on large volumes of recorded real-world behaviour.
The strategic logic is that a humanoid form factor needs no bespoke retooling of a workplace. A conventional factory robot is fast and precise but confined to a cell and a fixed routine. A humanoid, in principle, can be redeployed from one station to another the way a person can, which is why several manufacturers and software companies have converged on the format.
That promise remains unproven at scale. Humanoid robots are difficult across three fronts at once: balance and locomotion, dexterous manipulation with hands and fingers, and the general-purpose software that decides what to do next. Manipulation is the hardest of the three, because it involves contact, force and friction rather than the geometry problems that vision systems handle comparatively well.
Learning from human demonstration
The reason human workers enter the picture at all is the way these systems are trained. The dominant technique in contemporary robot learning is imitation: a model is shown many examples of a task being performed correctly and learns to reproduce the mapping from what the sensors see to what the joints should do. This is often called behavioural cloning, and it has become standard because writing explicit code for tasks such as folding cloth or seating a connector has proved extremely hard.
Imitation has a data problem. Text models were trained on a corpus that already existed on the internet; the equivalent corpus for physical manipulation does not exist. Video of people working is abundant but does not record the forces applied, the joint angles used or the corrections made when something slips. So the data has to be manufactured deliberately, and the usual methods involve people: teleoperation, in which a human drives the robot directly and the resulting trajectories are logged; and motion capture, in which a person wearing instrumented gear performs a task and the recording is mapped onto the robot’s body.
Both approaches make human physical skill the raw input. The specific equipment and protocol Tesla uses for Optimus data collection is not detailed in the material available here, and should not be assumed from general industry practice.
Worker reluctance to train a potential replacement
Against that background, the reported friction is straightforward. Ars Technica reports that Tesla employees asked to contribute to Optimus training have balked at doing so, understanding the work as helping to build a system that could displace them. The report frames this as a tension inside the programme rather than a formal dispute.
What is not known from the available material is how many workers have objected, whether the data-collection work was voluntary or assigned, what roles those workers hold, or how Tesla has responded. Those gaps matter, and filling them would require the primary reporting rather than inference.
The underlying situation is novel in one respect. Automation has historically been designed by engineers and installed over the objections, or with the negotiated consent, of the people affected. Imitation learning changes the position of the worker: the skill being automated has to be extracted from the person who holds it, and the transfer only works if they cooperate and perform the task well. That gives the demonstration step an unusual character — the labour supplied is the input to its own substitution — and it is the reason this particular friction has attracted attention beyond the company.
The target of 1,000 robots a week
Ars Technica reports that Tesla intends to reach production of 1,000 Optimus robots per week by the end of 2026. Annualised, that is on the order of 50,000 units a year, which would be a large number by the standards of the humanoid robot sector and a small one by the standards of car manufacturing.
Hitting such a rate is a manufacturing problem as much as a software one. A humanoid needs dozens of actuators, each with motors, gearing and sensing; the hands alone concentrate a great deal of mechanical complexity into a small volume. Components of this kind are not yet produced in automotive quantities anywhere, so a weekly output target implies building or securing a supply chain in parallel with finishing the product. Tesla has previously made public schedule commitments across its product lines that moved, and the material here does not establish current production volumes, so the target should be read as a stated intention rather than a forecast.
Capability is the separate question. Producing a thousand units a week is only useful if the software can perform economically valuable work reliably enough to be left alone — and reliability, not peak demonstrated ability, is the metric that governs deployment on a production line.
What the pieces add up to
Taken together, these items describe a programme whose bottleneck has moved. The hardware is being pushed towards volume manufacturing, and the stated constraint is no longer whether a humanoid can walk or grasp, but whether enough high-quality demonstration data can be gathered to make it broadly useful. That shift puts human skill at the centre of the timeline, which is exactly why a dispute about who provides the data is not a side issue.
It also clarifies what to watch. The production figure is checkable in time; the worker friction is the earlier signal, because it indicates that the cost of training data may not be purely technical. If demonstration work becomes contested — over consent, compensation or job security — then the data pipeline becomes an industrial-relations question as well as an engineering one, at every company pursuing this approach rather than only this one.
Sources and further reading
- Ars Technica — the technology publication whose report on Tesla worker reluctance and the stated 1,000-per-week Optimus target is the basis for the specific claims here.
- Academic robotics literature on imitation learning and behavioural cloning — for how demonstration data is collected and used to train manipulation policies.
- Tesla’s own public investor and product presentations — the primary record of what the company has stated about Optimus timelines and intended use.
- Labour research on automation and workplace technology — for the established framework on consent, bargaining and job displacement in automated production.
Surfaced from the rss:arstechnica signal “humanoid robot training dispute”. AI-assisted draft, editorially reviewed.
