Reusing Human Demonstrations Across Robots

Config is trying to turn robot data from a custom services cost into a reusable inventory asset. The key move is collecting demonstrations in human space first, then converting them into robot specific action traces later. That lets one sorting, packing, or assembly episode be sold as catalog data, reused in a Loop deployment, and adapted again for another arm or gripper, instead of paying to recollect the same task from scratch each time.1 2 3 4

• The cost side is labor heavy, so reuse matters. Config says its Hanoi studio has 100 plus operators producing about 20,000 hours of human data per month, and its model workflow starts with human actions because robot teleoperation is slower and more operationally expensive to scale.2 3
• This is different from vendors centered on one capture mode. Lightwheel sells simulation, human demonstration data, and evaluation tools. Maniformer adds wearable capture devices and simulation assets. DexForce leans into simulation, synthetic data, sensors, and robots. Config is more vertically organized around collect, convert, train, and deploy for bimanual tasks.5 6 7 8
• The underlying technical bet is real. Recent robotics research shows cross embodiment transfer can work when data is normalized into a common interface, with one framework reporting transfer across 16 embodiments and another showing average real world improvement from better cross embodiment data composition. That is what makes reuse economically meaningful rather than just a data storage claim.4 9 10

If Config keeps adding customers with overlapping manipulation jobs, its gross margin should improve as each new embodiment consumes more of the same underlying episode library. The company then starts to look less like a robotics contractor and more like a data platform whose best tasks get more valuable every time another robot can learn from them.1 3 4