Humanoid robots are showing up on factory floors faster than most manufacturers expected. What's less visible is how they're getting good at the job: Adapting to existing industrial environments based on real workers' skills and movements.
The old model was hand-programming: an engineer coded every motion for every task for every robot.
The model taking over instead is skill transfer: capture how a skilled worker actually performs a task, turn that demonstration into structured data, validate it in simulation, then deploy and reuse it: across tasks, across robots, and eventually across plants. It's the pattern behind nearly every serious industrial humanoid program running today.
The humanoid robot motion capture pipeline, end-to-end
Capture → retarget → simulate → test → deploy.
- Capture. A worker performs the task while wearing motion-capture hardware to collect data on joint angles and timing, data gloves for grip, and finger sequencing.
- Retarget. A robot isn't built like a person, and there are many industry-ready robots. They all have different limb ratios, different joint limits, and different actuation. Human motion has to be mapped onto the robot's build to make sense of the movement.
- Simulate. Before anything touches the line, the behavior gets stress-tested in simulation. This allows more variation, more edge cases in the safety of the digital environment.
- Test. The validated skill was transferred to the robot in a controlled pilot, with integration and safety teams involved from day one.
- Deploy. Once the robot skill is tested and validated through several iterations, the robot can start performing the task on the real workfloor.
Companies pioneering humanoid industrial automation
Automotive manufacturers, in particular, have become the proving ground for skill-transfer pipelines. Assembly plants were built for humans in the first place, which means the robot workers need to move in a human-like way to avoid the need to rebuild all the industrial plants.
- BMW has taken the most software-forward stance in public. Its Plant Landshut is developing the training data and simulation layer behind the company's humanoid work, explicitly building demonstration capture and policy learning as reusable infrastructure. The project uses Xsens Link and Metagloves by Manus, the hardware combination that is widely used across the humanoid-training industry. BMW has paired that software effort with real production hours: Figure's humanoid logged more than 1,250 operating hours and moved over 90,000 components at BMW's Spartanburg plant.
- Mercedes-Benz entered a commercial agreement with Apptronik to pilot the Apollo humanoid across its manufacturing facilities, targeting assembly-kit delivery and component inspection. The tasks chosen specifically because they're physically demanding and hard to staff.
- Hyundai, as Boston Dynamics' parent company, has its Atlas humanoid in active trials at its Georgia plant, using the same capture-retarget-simulate-deploy loop: motion captured from a human demonstrator, retargeted to Atlas's very different body, trained at scale in simulation, then brought back to the real robot.
- Amazon and GXO Logistics are running Agility Robotics' Digit in live warehouse operations, moving totes between conveyors: a logistics-side counterpart to the same underlying training approach, applied to a semi-structured rather than fixed-line environment.
Different robots, different companies, different tasks. The same underlying loop.
Why motion capture-based learning wins
Compared with hand-programming every motion, demonstration-based skill transfer offers:
- Faster task database creation. Record the movement; don't code it line by line, making constructing movement patterns faster.
- Better use of in-house process knowledge that would otherwise stay tacit, walking out the door with the worker who has it
- Access to unorthodox movement patterns. While a programmer can code the go-to movement trajectory, real operators can show their best tips and tricks for improved outcomes
- The real possibility of reusing one captured skill across sites and robot brands, rather than starting over each time
What this means if you're building the pipeline
Robotic industrial automation is underway, though it is still in its early stages. What will matter in the near future is the optimization of the skill transfer process: how easy it is to collect the data, how precise it is, how it connects to the robot of choice, and how this robot fits into the existing production environment. The manufacturers getting real results are the ones capturing breadth: multiple operators, multiple part variants, lighting changes, and, critically, failure recoveries, not just the clean successful run.
That's the layer Xsens Link is built for: Turning full-body motion into structured, reusable training data, the raw material every skill-transfer pipeline runs on.
Learn more about robot motion training with Xsens.