Config closes loop, Lightwheel supplies data

Config is trying to own the robot learning outcome, not just sell ingredients for someone else’s model stack. Lightwheel sells the world layer, the behavior layer, and the evaluation layer, meaning simulation assets, human demonstration data, and test environments. Config packages that same workflow into Loop, then uses it to train CFG-1 and deploy task policies, so a customer can buy a working manipulation model instead of assembling data, training, and validation on its own.1 2 3 4

• Lightwheel’s product menu is modular. SimReady Library supplies simulation assets and scenes, EgoSuite supplies egocentric human demonstrations, and RoboFinals evaluates robot models. That fits labs that already have model teams and want better training inputs and tests, not an outside model provider.1
• Config’s position is more integrated. Existing research describes Loop as spanning collection, conversion, training, deployment, and evaluation, with CFG-1 as its in house bimanual model. It also supports outside models like NVIDIA GR00T and Physical Intelligence Pi 0.5, which lets it sell both full service outcomes and infrastructure.2 3
• Maniformer sits closer to the data supplier side of the market. Its MEgo hardware captures first person and gripper data, and its site emphasizes data acquisition, governance, and crowdsourced collection. That makes the market split clearer, some vendors sell picks and shovels, while others also sell the robot brain.4 5 6

The category is moving toward closed loop suppliers that can prove data quality by turning it into deployed robot behavior. Infrastructure only vendors will keep winning with customers that want model control in house. Integrated vendors will win where buyers care more about faster task deployment than owning the full training stack.1 2 3 4 5 6