Tesla Optimus New Progress Revealed: Data Collection Relies on Hiring People to Do Housework, Motion Capture May Be Abandoned

High-end robot data often uses the most basic collection methods.

 

According to Business Insider, Tesla is using a data collection team to train the Optimus robot to move like a human. In 8-hour shifts, data collectors repeat the same actions hundreds of times, like picking up cups, wiping tables, pulling curtains. Before starting, they get task assignment files and operation manuals to make sure the data collection is done right. Each employee needs to collect at least 4 hours of usable video per shift.

 

Notably, Tesla is now collecting data through cameras instead of motion capture suits or manual controls like before. In June this year, after Optimus project head Milan Kovac left, the company told the data collection team: "Without motion capture suits, data collection can scale up more."

 

The work is physically demanding. Besides surrounding cameras, each collector wears about 5 cameras on a helmet and heavy backpack for full recording. Insiders say at Tesla's Fremont factory in California, collectors wear headphones and backpacks, sorting vehicle parts and working on conveyor belts.

 

 

The long-term physical labor takes a toll. Some employees say uneven backpack weight caused back and neck injuries during shifts. Others note collectors have gotten motion sickness from wearing head-mounted devices for long periods.

 

In terms of team size, the report says at its peak, Tesla had over 100 employees doing data collection. Behind this is Musk's humanoid robot ambition: In the Q3 earnings call this year, Musk claimed the company would hit 1 million Optimus units annually. He also said humanoid robot business would account for about 80% of Tesla's value in the future.

 

Data has always been seen as key to improving humanoid robots' generalization. Huachuang Securities points out that fusing multimodal training datasets will greatly boost robots' environmental perception and multi-task handling. By type, data splits into real and simulated. Real data is considered the best "gold data" for training but costs more and lacks uniform format. Simulated data can be generated at scale cheaply, but trained models often don't adapt well.

 

In this context, "real-virtual combination" is the mainstream data collection approach in robotics now. Liu Yufei, deputy GM of the national-local joint humanoid robot innovation center, says they've launched real-virtual training fields in 8 provinces and cities. IDC China research manager Li Junlan predicts the industry will build a solid data foundation based on massive high-fidelity physical data plus high-quality real collection, quickly improving intelligent generalization.

 

According to Research Nester's September report "Data Acquisition System Market Size and Forecast," the market will exceed $2.4 billion in 2025, reaching $3.98 billion by 2035, with a CAGR of about 5.2% over the forecast period (2026-2035). In 2026, the industry size is estimated at $2.51 billion.

 

But at the same time, signs suggest future robot training could become "AI-ized." For example, Tesla recently announced training Optimus in its self-developed world model. Dongwu Securities says companies currently use world models, teleoperation, simulation migration, simulated training, etc., but all have limits and can't achieve general generalization. Embodied intelligence learning methods still need exploration.