Conversion Engine for Robot Retargeting

This is the part of the stack that lets Config sell one human demonstration pipeline across many robot bodies. A person can show the task with human hands, but the deployed robot may have different finger widths, joint limits, reach, and grip points. The conversion layer turns that mismatch into usable training data, which means customers do not need to recollect every task on every arm and gripper combination before training starts.1 2 3 4 5

• In practice, the hard part is not recording motion, it is retargeting it. Research on human to robot dexterous manipulation shows that human joint motion cannot be copied directly across different embodiments, so systems preserve object contact patterns and feasible end states, then solve for robot specific trajectories that fit the target hand or arm.3 4 5
• That makes Config different from data only suppliers such as Lightwheel and Maniformer. Those companies sell simulation, capture hardware, datasets, or evaluation tools. Config bundles collection, conversion, model fine tuning, and deployment, so the customer buys a working task policy rather than just raw demonstrations.1 6 7
• It also explains why Loop supports outside models like NVIDIA GR00T and Physical Intelligence Pi 0.5. Once demonstrations are converted into robot ready action data, Config can route that data into its own CFG-1 or third party vision language action models, which widens the set of robots and customers it can serve.1 2 8 9

The next advantage will come from how many embodiments this conversion layer can cover with little extra setup. If Config can keep turning the same human workflow into reliable trajectories for new arms, grippers, and bimanual systems, it becomes less a labeling vendor and more the translation layer that sits between human skill and robot deployment.1 2 3 8