Build vs Buy Robot Learning

This choice determines whether robotics teams are buying a toolchain or outsourcing the hard part of robot learning. Lightwheel and Maniformer sell pieces of the pipeline, like simulation assets, capture hardware, or converted demonstrations, for customers that still want to train their own policies. Config goes further by pairing its Loop workflow stack with CFG-1, so a buyer can hand over data collection, task tuning, and deployment instead of staffing that capability in house.1 2 3 4 5

• Lightwheel is closer to infrastructure. It sells simulation and data systems that help customers make and test training examples, which fits teams that already have model engineers and want control over how policies are built.1 2
• Maniformer sits in the middle. It offers capture hardware, a task catalog, and converted human demonstrations, but it is positioned more as a data supplier than as the seller of a full robot brain, which keeps more training work on the customer side.3 5
• NVIDIA and Physical Intelligence push the market toward insourcing. NVIDIA packages open GR00T models with an end to end workflow for capture, post training, evaluation, and deployment, while Physical Intelligence has published open robot foundation model work, giving buyers starting models they can adapt with their own data teams.4 6 7

The market is heading toward a split between horizontal model stacks and full service task suppliers. As open model ecosystems improve, more well funded robotics teams will assemble their own pipelines around GR00T like or PI style models. That makes turnkey vendors strongest where customers care less about owning the training stack and more about getting a task working fast on real hardware.4 6 7