ByteDance has opened dozens of humanoid robotics roles inside its Seed Model division. The listings were refreshed on 27 October 2026, and their content marks a shift — from backing other companies' robots to building its own hardware. One listing states the goal plainly: carry a prototype from the demo stage through to stable, series delivery.
Key takeaways
- ByteDance's official careers search returns 55 listings under the term 具身智能, or "embodied intelligence"
- Seed Model is hiring a gearbox designer, a body structure engineer, a motor control engineer and a wiring harness engineer, all in Beijing
- Several listings carry a named sub-team: Seed Robotics
- The hardware product manager listing names the EVT, DVT and PVT production gates
- ByteDance already runs more than a thousand logistics robots in Douyin e-commerce warehouses
Where this starts
A Douyin e-commerce warehouse. A mobile robot shuttles cartons between the sorting zone and the loading bay. It works because the task is narrow — flat floor, known parcel dimensions, a repeatable route.
Now a humanoid is meant to do the same job. Grip an irregularly shaped parcel, move it from shelf to trolley, and repeat that ten thousand times.
The obstacle is not the algorithm. It is the gearbox in the shoulder that develops backlash?backlash: excessive free play between gear teeth, which stops the joint position from matching the commanded value after two weeks of work.
And that is why a company known for apps is now hiring engineers who specialise in gear tooth profiles.
The warehouse scenario is an editorial simplification. The figure of more than a thousand logistics robots in Douyin warehouses comes from TMTPost's reporting.
What exactly did ByteDance open?
The listings sit on the company's official careers board. Searching 具身智能 returns 55 entries — some are conventional AI roles covering VLA models, world models and reinforcement learning.
What sits next to them is more interesting.
Seed Model is hiring a robot gearbox design engineer in Beijing. The requirements are mechanical, not software: three years on planetary gearboxes, tooth profile optimisation, heat treatment such as carburising and nitriding?carburising and nitriding: thermochemical processes that harden the surface layer of a gear tooth to improve its wear resistance, plus gear grinding and hobbing?hobbing: a method of cutting gear teeth with a rotating worm-shaped cutter, the standard process in gear manufacturing processes.
Alongside it the company wants a body structure engineer to design joint modules, including bearing and material selection. Then a motor control algorithm engineer, an embedded driver engineer, a wiring harness engineer and a physics simulation engineer.
Several titles carry a distinct name — Seed Robotics. The Seed team was formed in 2023 and is responsible for the Doubao model among others, with robotics listed in its official description beside MLLM?MLLM: a multimodal large language model, processing images, audio or video alongside text, media generation and AI for Science.
What does the road from demo to production look like?
The clearest statement of intent sits in the listing for a "mass-production class" hardware product manager, where the responsibilities are laid out in stages.
This is a simplified sequence built from the responsibilities in the listing, not a description of a working ByteDance production line.
What do these names mean?
A handful of terms recur in the listings, and the task lists make little sense without them.
- Embodied intelligence — AI that drives a physical body operating in the real world. Unlike a cloud model, it pays for mistakes in broken hardware.
- Planetary gearbox — a gear assembly that cuts motor speed and raises torque. One sits in every robot joint. The equivalent of a gearbox in a car.
- EVT, DVT, PVT — three successive prototype rounds before mass production. They ask, in order: does it work, can it be built repeatably, can the factory hit its yield. The equivalent of alpha, beta and release candidate.
- **Sim-to-Real** — transferring a skill learned in a simulator onto real hardware. Simulation allows a million attempts a day, which no workshop floor can match.
- VLA — a model that turns a camera image and a text instruction directly into motion, instead of chaining separate perception, planning and control modules.
Together these terms describe one loop. A model trains in simulation, lands on hardware, and that hardware has to clear the production gates and survive daily use.
Why not just buy a finished robot?
It is the obvious question to ask. Chinese manufacturers have ready-made platforms, and a full humanoid component set can be sourced in the Yangtze and Pearl River deltas. ByteDance already took a stake in Unitree in June 2025 through the affiliated Jinqiu fund — as a financial investor at the time.
The trouble is that a bought robot closes off change exactly where the advantage is decided.
A model trained in simulation only works if the simulated joint dynamics match the real ones. When the gearbox comes from an outside supplier, its friction, backlash and wear are parameters that can be neither measured precisely nor tuned to the model. That is the classic Sim-to-Real gap.
There is a second, less technical reason. A humanoid that fails in scenarios outside the test script will not enter any warehouse — and liability for a breakdown falls on whoever supplies the whole system, not on the gearbox maker. Controlling the hardware is therefore the price of accepting that liability.
Is it working yet?
For now this is recruitment, not a product. ByteDance has announced no humanoid, no specification and no date. Only the Seed Model job listings and the Seed Robotics team name are publicly confirmed.
An earlier, simpler line is in production, though. According to TMTPost, ByteDance's logistics robots have passed a thousand units and handle sorting and transport in Douyin e-commerce warehouses, with SF Express and BYD Electronics among external customers. These are wheeled machines, not humanoids.
A year earlier, in November 2025, the Volcano Engine cloud platform sought a senior embodied intelligence expert at 95,000 to 120,000 yuan a month, roughly …16,500. The current wave of listings is visibly broader and reaches into mechanics.
Why does this matter?
Chinese internet firms have so far entered robotics through capital and models. The job list inside Seed Model points somewhere else — at taking control of the layer where a humanoid's cost and durability are actually settled. If advantage in this segment rests on gearboxes and joint service life, then an algorithmic edge cannot be rented along with someone else's chassis.
What next?
- The listings reference the EVT, DVT and PVT gates, so the first verifiable signal will be an engineering prototype rather than a video
- Hiring for the gearbox designer and body structure engineer is running in Beijing — filling those roles will show whether ByteDance builds its own joint or integrates someone else's
- Rivals are moving in parallel: Unitree showed the full-sized H2-X humanoid this week, narrowing the window to join the field
Sources
- TMTPost — 字节跳动招人，入局具身智能
- ChinaTechNews — ByteDance Enters Humanoid Robotics Race as Unitree Debuts Advanced Model in China Tech Surge
- ByteDance Careers — 机器人减速机设计工程师 - Seed Model
- TechNode — ByteDance launches high-paying recruitment for humanoid robotics expert
