Tesla ramps Optimus to a whole bunch per week, however the robots cannot generalize

Tesla is now constructing a number of hundred Optimus humanoid robots per week at Fremont, roughly 10 instances its Q2 output, in response to a brand new report. But the robots coming off the road nonetheless can’t deal with generalized duties – limiting their usefulness.

The report additionally describes fragile palms, misaligned elements on the meeting line, and suppliers that may’t maintain high quality at quantity. Those are the 2 issues that determine whether or not humanoid robots make cash: generalization and reliability.

Hundreds of robots, largely for Tesla

The Information reported this morning (paywalled) that Tesla went from just a few dozen Optimus items per week throughout small-batch testing in Q2 to a number of hundred per week in August. Managers are reportedly aiming for a steady automated line able to greater than 1,000 robots per week by the tip of the 12 months, with an eventual goal of about 20,000 per week.

That line lives the place the Model S and Model X was once constructed. Tesla ended manufacturing of its two flagship automobiles in early May and converted the Fremont line to Optimus, pulling employees and engineers from S/X, and dozens extra from Model Y, onto the robotic program.

Advertisement – scroll for extra content material

So the place are all these robots going? Mostly nowhere. According to the report, most items are used internally for testing, coaching, and information assortment. The ones working inside Tesla’s factories are confined to tightly managed, supervised areas, they usually’re programmed for particular duties somewhat than working as general-purpose machines.

And the V3 robots Tesla is constructing now aren’t even the model it plans to commercialize. That one nonetheless has to cross stricter sturdiness and reliability thresholds.

The palms are the arduous half

Musk has promised human-level dexterity for Optimus. Getting there at scale is proving painful.

Per The Information, the hand and forearm include over 100 screws and small parts that employees nonetheless assemble by hand. Fixtures on a number of stations, together with palms, joints, and electronics testing, can’t persistently line up elements which can be far smaller and tighter-tolerance than something on a automobile. The result’s extra robots needing rework after coming off the road.

Durability is its personal downside. Some of the hand’s contact sensors have had reliability points, and Tesla’s repair is a replaceable “sensing glove” that it plans so as to add subsequent 12 months so it doesn’t should swap the entire hand.

Then there’s the availability chain. Tesla depends on exterior suppliers, many in China, for motors and the precision gears that drive the joints. Some of them could make good elements in prototype portions however wrestle to maintain high quality constant at larger quantity.

Days to be taught a primary activity

The larger problem is the mind. Three folks conversant in the system informed The Information that Optimus’ AI can’t but reliably deal with a variety of duties, and that its conduct could be unpredictable in conditions it hasn’t been skilled on.

Tesla is making an attempt to construct a library of primary actions the robotic can recombine for brand new jobs. But it nonetheless reportedly takes a number of days for Optimus to be taught even primary duties.

To feed the fashions, Tesla has over 500,000 hours of coaching information and needs to double that by year-end. It has moved a lot of its self-driving information annotation staff to Optimus, employed devoted information collectors carrying digital camera helmets and motion-capture fits, and arrange coaching hubs in Colorado, Arizona, and Florida.

The plan for patrons follows the identical logic. Tesla reportedly intends to lease, not promote, Optimus to a brief record of corporations whose factories and warehouses appear to be its personal, so the robots have a better time adapting, and to make use of information from these deployments to enhance the AI. That’s the FSD playbook: ship the {hardware}, gather the fleet information, and promise the software program will catch up.

This can also be a far cry from what Musk was saying final 12 months. In January 2025, he mentioned Tesla would construct about 10,000 Optimus robots in 2025 and that “a number of thousand” could be doing helpful work by year-end. A 12 months later, he admitted that zero Optimus robots were doing useful work at Tesla. He additionally promised a V3 reveal by mid-2026. It’s late September, and that demo nonetheless hasn’t occurred.

And opponents aren’t ready. XPeng started running an automated IRON humanoid production line in Guangzhou this month, focusing on business gross sales in 2027.

Electrek’s Take

We assume humanoid robots will change into a giant new class of robots. But we additionally assume they’ll be a small fraction of the general robotics business. For most jobs, a robotic designed for that particular job, like a robotic arm on a line or a wheeled warehouse bot, will probably be a number of instances extra environment friendly than a robotic formed like a human. The humanoid’s solely actual benefit is that it could theoretically do something in an area constructed for folks.

That benefit solely exists if two issues occur. First, the AI must generalize. A humanoid that must be skilled for days on every activity and stored in a fenced-off space is simply an costly, much less environment friendly specialised robotic. Nobody has cracked generalization but, and it’s extraordinarily arduous to foretell when anybody will. Second, the robotic must be dependable sufficient over years of use to pay for itself. Touch sensors failing and 100+ hand-assembled elements per arm aren’t a fantastic begin.

Based on this report, Tesla has issues with each.

But it wouldn’t be the primary time Tesla made massive product selections based mostly on AI breakthroughs that haven’t occurred but. In 2016, it began telling clients each automobile it constructed had the {hardware} for full self-driving. It has since admitted HW3 automobiles will want an improve, and “Full Self-Driving” nonetheless requires supervision practically a decade later. Now Tesla has killed the Model S and Model X to construct a robotic whose AI can’t do what the product wants it to do.

Maybe Tesla will get there. But the corporate is once more betting its manufacturing facility traces on a software program timeline it could’t predict. How many instances does that have to occur earlier than traders begin pricing it in?

If you’re following the AI and robotics increase, you’ve most likely observed what it’s doing to electrical energy demand and charges. With electrical energy charges up nearly 10% final 12 months and anticipated to maintain climbing, going photo voltaic is without doubt one of the finest methods to guard your self in opposition to rising prices. And with lease and PPA choices, you are able to do it with zero upfront price and begin saving instantly. If you need to discover the very best deal, take a look at EnergySage. It’s a free service with a whole bunch of pre-vetted installers competing to your industry, so that you save 20 to 30% in comparison with going it alone. No gross sales calls till you decide an installer. Get your free quotes here.