On Thursday 3 September 2026, CJ Logistics said it had placed two dual-arm humanoid robots into the packing process at the Yangji Olive Young fulfillment center in Yongin, Gyeonggi Province. Read the claim precisely: not a demonstration or proof of concept, but a live operating process handling real customer orders. The robots put cushioning material into shipping boxes. A worker then loads the ordered goods, and the box ships to an actual customer. Kim Jeong-hee, who heads CJ Logistics’ TES Logistics Technology Research Institute, framed it as moving past “showing technology” into “performing a role in real logistics operations.” The distinction is real. It is also two machines on one station.
The platform is not a mystery. Edaily named it as ROBOTIS (KOSDAQ 108490) “AI Worker,” two units, the same platform as the earlier trial. In September 2025 CJ put a ROBOTIS humanoid on the Gunpo packing line to check efficiency and safety on cushioning replenishment, under a physical-AI agreement signed with ROBOTIS chief executive Kim Byoung-soo. CJ said then it would finish validation by end of 2025 and begin phased rollout the year after. That schedule roughly held, which is more than most humanoid timelines manage.
Twenty-five joints, and one of them is a finger
ROBOTIS ships the AI Worker in FFW-SG2 mobile and FFW-BG2 stationary trims. CJ printed no trim or joint map for the Yangji machines, so treat the published sheet as the closest hardware page, not a confirmed build. On the FFW-SG2: 25 degrees of freedom, split seven per arm, one per gripper, two in the head, one in the lift column, six in the base. It is 604 × 602 × 1,623 mm at 90 kg. Arm reach is 641 mm to the wrist, before the hand. Nominal payload is 3.0 kg per arm and 6.0 kg dual-arm, peaking at 5.0 and 10.0 kg. The base is a swerve drive rated 1.5 m/s. Power is a 25 V, 80 Ah pack, so 2,040 Wh. Compute is an NVIDIA Jetson AGX Orin 32 GB. Sensing is three RGBD cameras, two LiDAR units and an IMU. Internal comms are RS-485 at 4 Mbps. Arm joints one through six are DYNAMIXEL-Y, joint seven is DYNAMIXEL-P, the head runs DYNAMIXEL-X, and the lift is DYNAMIXEL-Y with 0 to 500 mm of stroke. Ambient rating is 0 to 40 °C. The stationary FFW-BG2 drops to 19 DoF and runs off wall power.
Now count the wrists, because that is where the story sits. Each gripper is a single degree of freedom. The standard end effector is ROBOTIS’s RH-P12-RN, a two-finger unit, and the dexterous finger actuator is listed on the same sheet as still in development. Edaily reported the Yangji units run exactly that, a two-finger gripper in place of a human-like hand, deliberately, because the process is standardized. FIRGELLI does not read that as a shortcut. It is the right call for the task chosen, and it tells you which task was available.
The job is the easy half of packing
Cushioning insert is close to the friendliest manipulation problem in a fulfillment center. The material is paper or air pillow, so grasp pose is forgiving, placement tolerance is wide, and nothing is fragile. The box arrives square, open and fixtured. Edaily adds that the robots also close the lid, which is an edge-alignment problem rather than a grasp problem. A person still loads the products. The humanoid is not standing in for a picker; it absorbs two motions bracketing the human’s.
Put the payload figures against the task and the mismatch is plain. Three kilograms nominal per arm, and the object is void fill weighing tens of grams. Torque is nowhere near binding. The constraints are perception on a deformable object, cycle repeatability, and lid registration. That is the honest reason this task went first. CJ’s own roadmap concedes the ordering: cushioning now, then picking, sorting, inspection and packaging, and eventually one machine running several processes back to back as a general-purpose logistics humanoid.
None of the numbers that decide whether the station pays are public: cycle time per box against a trained human, uptime across a shift, or boxes per operator touch. Two units and no takt figure is a pilot wearing production clothes, and Edaily called it a trial run still needing verification on speed, stability and cost.
The hand that unlocks picking is a 2028 program
Picking mixed goods is where a logistics humanoid would earn money, and picking needs a hand. CJ’s partner there is Aidin Robotics, which leads a Korean national R&D project for a logistics humanoid carrying a high-sensitivity robot hand on a multimodal AI foundation model. It runs to 2028 at KRW 5.1 billion total, including roughly KRW 4.1 billion of government funding, with CJ Logistics, the Korea Electronics Technology Institute and Sungkyunkwan University participating.
The hardware is the AIDIN-Hand. Generation one is 15 DoF with a 15 kg payload. Gen2, shown at AW 2026 in Seoul, adds a thumb degree of freedom for 16 total and lifts payload to 20 kg, with a tactile array across the front surface and a six-axis force/torque sensor at the fingertip, on a linkage-driven mechanism. Aidin still lists the hand as research-purpose, with Gen2 release scheduled for the second quarter of 2026.
So: a one-DoF pincer placing void fill today, versus the 16-DoF sensorized hand that would make picking real, on a 2028 horizon. That gap is the schedule worth reading.
The brain is a partnership, and it learned from teleop video
CJ is not writing the policy stack itself. The robot foundation model work runs through RLWRLD, a physical-AI company led by Ryu Jung-hee that CJ signed an MOU with and backed in a Seed2 round. RLWRLD says its RLDX model fine-tunes onto any body from a precision hand to a full humanoid, and that it deliberately skips whole-body walking to concentrate on fingers, grasping and contact-aware control at industrial speed, trained on 4D+ motion capture from real factory floors rather than simulation.
The Gunpo training method explains the capability ceiling. Operators teleoperated the robot through the cushioning task, the runs were recorded to video, and the model learned by watching them repeatedly. CJ now cycles live Yangji data back into training, blended with simulation data. Imitation from teleoperated demonstrations reliably gets one motion working on one station. It is also why “expand to picking” is not a software update: new task, new demonstrations, new failure modes, and an end effector that cannot form the grasps the new task needs.
What a buyer should ask next
Keep the questions boring and hardware-shaped. Which trim is at Yangji, and if it is the mobile one, does the swerve base drive during a shift or sit parked. Cycle time per box against the human baseline. Runtime on the 2,040 Wh pack at real duty, and whether packs hot-swap between shifts. Case temperature at shoulder and elbow after a full shift, against that 0 to 40 °C ambient rating. Boxes per operator intervention. Service interval on the RH-P12-RN under abrasive void fill, and whether that gripper suits lid closing at all. And the one governing the roadmap: the committed swap date for a sensorized hand.
FIRGELLI’s read on 3 September is that CJ Logistics did something modest and honest. Two ROBOTIS AI Workers, a one-DoF two-finger gripper, void fill and a lid, real customer boxes, on a line where a human still loads the goods. The hardware page underneath is unusually specific for this sector: 25 joints, DYNAMIXEL actuation throughout, 641 mm reach, 3 kg nominal per arm, a 2,040 Wh pack. Detail like that beats a choreography video. Missing is every figure that turns a station into a business case: takt, uptime, intervention rate, thermal behavior across a shift. The closest architecture we track is Rainbow Robotics RB-Y1: wheeled base, lift column, dual arms, different vendor. If you stack this against the factory sequencing pilots or food-service cells covered on this desk, hold it to the same bar: a named trim, a published cycle time, a measured wear number under real duty. September put two arms on a packing line. It did not publish the shift.