RLWRLD and CJ Logistics expanded their collaboration with a memorandum of understanding to co-develop logistics-specific robotics foundation models and move them from lab checkpoints into live warehouses, with Korea as the first proving ground and overseas commercialization as the stated next step. The division of labor is explicit. CJ Logistics supplies operating data from express, fulfillment and cold-chain sites, defines process requirements, and validates performance against real throughput, exception and safety criteria.
RLWRLD trains the foundation models and writes the control software that turns multi-modal perception into motor commands. The core model family is RLDX-1, described as a dexterity-first foundation model that fuses vision, language, torque, tactile signals and task memory rather than treating warehouse work as a vision-language-action problem alone. That distinction matters in logistics: cartons crush, labels sit at odd angles, dunnage shifts, and the same SKU can arrive in multiple pack types.
First field work includes parcel-sort validation in which robots must align shipping labels and capture grasp traces for further training. CJ already runs dual-arm humanoids at the CJ Olive Young center in Yangji, inserting cushioning into outbound cartons—an early live-line use of humanoids in a commercial logistics building rather than a demo cell. The partners intend to extend the same stack to picking, sortation and inspection, with a longer-term target of a general-purpose logistics humanoid.
After proof-of-concept, they plan a Logistics-as-a-Service wrapper covering order intake, storage, pick-pack, dispatch, returns and analytics, so the software brain can be sold independently of a single robot body. The strategic bet is that warehouse data, not only simulation, is the scarce asset for physical AI: messy live exceptions produce training signals that scripted cells never generate. For the industry, the story is less a single robot launch and more an operator-plus-model-lab alliance trying to industrialize foundation models where SKU variety and shift-to-shift layout change punish brittle scripts.