Apple's humanoid robot hits 91% success rate as Tesla's Optimus shifts to rent-only

Highlights

• Apple's HALO runs onboard inference on a Unitree G1 plus Mac mini, hitting 91% success on single tasks
• Long-sequence compound tasks drop to 70%, with the most complex combination at just 50%
• Musk: Optimus may be rent-only, with initial cost around $30,000
• Humanoid robot daily rental prices fall below 1,000 yuan, yet platforms keep piling in

Apple quantifies the bar for "can actually work"

Apple's research team submitted the paper "Perceptive Humanoid Motion from Simulation" to arXiv on Nov 4, disclosing HALO, a humanoid robot autonomous control system, and releasing code for data generation, model training and deployment. The hardware is a Unitree G1 fitted with a 3D-printed back shell housing a Mac mini (M4 Pro, 64GB unified memory) for onboard real-time inference, while the G1's onboard chip handles high-frequency motion tracking. (Source: IT Home)

The system has three layers: a top-level task orchestration module built on a vision-language model, defaulting to Qwen3.6-35B-A3B; a mid-level motion generation layer using Qwen3.5-2B; and a bottom-level motion tracker trained via reinforcement learning and policy distillation, running 50 times per second. Training data was synthesized with 250 Nvidia H100 GPUs over about 48 hours, and the motion generator was trained on 64 B200 GPUs. (Source: IT Home)

The key numbers are the tiered success rates: 360 single tasks across six real-world environments with no prior scanning or mapping yielded an average 91% success rate. But long-sequence compound tasks, each tested 10 times in a single real-world scenario, averaged 70% across three combinations—and the most complex one, "duck under an obstacle—walk to the target—push the box," succeeded only 5 times out of 10. The team acknowledges the robot relies mainly on vision and struggles to directly sense contact; performance drops on tasks involving more contact when objects leave the field of view or are occluded, and adding tactile sensing is the next step. (Source: IT Home)

91% means single actions are usable; 50% means "working continuously" remains the dividing line. Simulation data hasn't fully escaped manual work either: new physical tasks still require humans to write prompts and design keyframes.

Tesla: Optimus pivots to rental, capacity planning comes first

Speaking at the Baron Investment Conference, Musk said Optimus may adopt a "rent-only, no sale" model, with rental pricing and the work it can perform not yet disclosed. Initial per-unit cost is estimated at about $30,000, later expected to fall to around $20,000. Tesla began early Optimus production at its Fremont factory in October, using space freed up from the former Model S and Model X lines, with planned annual capacity of 1 million units. But Musk expects at least another year to reach full production speed; the plan is to deploy them in its own factories first, deliver to external companies around late 2027, and launch the AI5-equipped version around April 2027. He also claimed its production efficiency will reach five times that of an average human. (Source: IT Home)

At the same CIIE, Tesla showed the gold, steering-wheel-free, pedal-free production Cybercab in its Asia debut. The vehicle began operating in Austin, US, on Sept 3, with more than 300 vehicles in service within 30 days; as of Oct 18, the Robotaxi fleet had logged about 5 million kilometers with no safety monitor in the vehicle, zero airbag deployments and no active collisions injuring anyone. Musk also said Tesla will dismantle the Cybertruck line around June-July next year to expand Cybercab capacity, adding a production shift to clear backlogged orders. (Source: IT Home)

China signals: hot buzzwords, cold deployment

The China Science Writers Association released its top 10 science popularization buzzwords for 2026, with AI-related terms taking seven spots—including "humanoid robot," "embodied intelligence," "human-machine collaboration" and "brain-computer interface"—reflecting still-rising public awareness. (Source: IT Home)

But the industry side shows cooling signs. Two TMTPost reports published the same day pointed to the same thing: one said humanoid robots that "don't want to work" are gathering dust in warehouses—the question isn't how high a robot can flip, but how far it is from the workshop floor; the other said robot daily rental prices have fallen below 1,000 yuan while platforms keep entering the market. Falling rental prices and platforms doubling down at the same time mean the industry is removing the rental "cushion" and shifting to more direct competition on cost of use. (Source: TMTPost)

Put the three threads together: on the technology side, success rates are now disclosed by task complexity; on the commercial side, the model is shifting from selling hardware to renting services; and rental prices are already falling. What's worth tracking next is whether tactile sensing can close the gap on long-sequence tasks, and when Optimus rental pricing will be made public—those two numbers will directly determine whether the "entering the workshop" math adds up.