RLWRLD robot dexterity research could help with a specific warehouse task. A parcel needs to arrive with its shipping label facing up. The hand may have to lift, turn or reposition the package while the next station waits. In September 2026, CJ Logistics named that sorting step as an example for a planned field trial with RLWRLD.
A related movement appears in RLWRLD’s May technical report. On a physical ALLEX robot, its RLDX-1 model picked a box from a moving belt and placed it on a shelf. The researchers tested four speeds. At the two speeds used in training, all the reported trials succeeded. At the two intermediate speeds, the result was lower.
The two records point to different stages of work. One documents a physical experiment with a stated protocol. The other sets out joint development and a proposed trial in a logistics operation. AWS also describes the infrastructure that supports the model’s training. The sources were checked on November 13, 2026.
CJ Logistics announced the new agreement on September 21, 2026, after a signing on September 18 in Seoul. The companies plan to develop a robotics foundation model for logistics and pursue proof-of-concept testing in operating facilities. It also follows a November 2025 agreement and CJ Logistics’ participation in an equity investment. CJ Logistics’ September 21 announcement
Under the plan, CJ Logistics brings operational data and facilities. It will help set and verify the automation requirements and performance criteria. RLWRLD will lead model and physical AI software development. The release also points to hub and sub terminals as possible settings and names label-up parcel orientation as an example.
That sorting step helps explain why warehouse robotics teams care about dexterity. Turning a parcel can require choosing where to grip, maintaining control as it rotates and setting it down in a usable position. For example, a carton and a flexible mailer can respond differently to the same movement.
The announcement describes the work the partners intend to pursue. Its practical value for a reader is the connection between a proposed task and the parties responsible for it. A subsequent trial record would identify the chosen equipment, site and achieved performance for that work.
An April 2025 investment announcement by ANA Holdings identifies RLWRLD’s location as Seoul and dates its founding to July 2024. It names Jung-hee Ryu as chief executive. CJ Logistics refers to the company as 리얼월드 in Korean and places the September signing at its Seoul premises. ANA Holdings’ investment announcement
RLWRLD’s current contact page lists a US headquarters in San Francisco and offices in Seoul and Tokyo. The dated investor announcement and current office listing describe different moments in the company’s development. Together, they place the Korean operations in a company that also has a presence in the United States and Japan. RLWRLD’s office listing
For readers following physical AI in Korea, RLDX-1 is the model behind this story. RLWRLD’s product focus is robotic manipulation software. Its technical report lists authors affiliated with RLWRLD and KAIST and evaluates the model with several robot platforms.
Seoulz previously examined RLWRLD within the country’s wider physical AI industry. That coverage introduces the company and its approach to training data. The later CJ announcement then gives the warehouse story a particular handling task to examine. Seoulz’s coverage of Korea’s physical AI sector
A vision-language-action model converts observations and instructions into robot actions. RLWRLD calls its architecture the Multi-Stream Action Transformer. It uses separate streams for different inputs and lets them interact. The company describes work with visual information, instructions, movement history and physical signals. RLWRLD’s RLDX-1 technical blog
The hardware determines which of those signals are available. A camera can show where an object is. Joint torque can provide information about the load as a robot lifts or turns it. Meanwhile, tactile sensors on a different setup can provide measurements at contact. Each input describes a different aspect of the movement.
In the report’s ALLEX experiment, the setup used a head-mounted stereo camera and joint torques estimated from motor currents. The physical platform had two arms and two five-finger hands. Those details matter because changing the hardware can change the observations and available movements. For example, a sensor at the hand and a torque estimate at the joint measure different physical effects.
Consider what happens when a parcel partly disappears behind a robot’s fingers. A camera may show its visible edge, while the motion and load signals give additional information. A study of robotic manipulation can therefore ask which inputs a specific task uses and how performance changes as the conditions vary.
On ALLEX, the researchers asked the robot to pick a box from a moving conveyor and place it on a shelf. Two of the four speeds appeared during training. The other two were intermediate speeds, so the test included movement conditions the model had encountered and two it had yet to encounter. The report’s appendix specifies six trials per speed and 24 trials in total. RLDX-1 Technical Report, Section 6.3 and Appendix G.3
The authors report 100 percent success for the training-speed group and 75 percent for the intermediate group. With 12 trials per group, those rates correspond to 12 and nine successful trials respectively. Adding them gives 21 of 24, or 87.5 percent, for that test.
The appendix also describes task-specific fine-tuning. For each task, the researchers enabled the module relevant to the tested capability. The conveyor result therefore describes this model preparation, physical robot and task. Consequently, the intermediate speed comparison provides evidence of adaptation within the tested range.
Warehouse robotics would put a different set of demands on a handling system. A conveyor can carry items with other shapes, surfaces and starting orientations. The next test can specify which of those conditions it changes. A later comparison becomes easier to interpret when the reader can match the new trial to the original protocol.
RLDX-1’s technical report gives an ALLEX aggregate of 86.8. The company’s table lists four task values on the same displayed scale. Conveyor handling scores completion, and the hidden-object task scores selection of the correct box. The card and pouring tasks award progress credit as intermediate stages are completed.
For the card task, the stages are sliding, lifting and handing over the card. The pouring test scores stages from grasping the objects through presenting the cup. That protocol gives partial credit when an attempt advances partway. Consequently, the reported average combines completion or selection rates with progress scores.
Both scoring approaches can tell a researcher something about robotic manipulation. Binary scoring answers whether an attempt reached its defined endpoint. A progress measure also identifies how far the attempt got. For example, a card may reach the table edge before a grasp fails.
In parcel handling, that distinction could guide the next engineering decision. Partial progress could identify a difficult grasp or turn. A completed placement would have its own operational definition, including the orientation needed by the next station. The two figures would describe different aspects of the same trial.
AWS’s October 28 partner account cites 58.7 percent for RLDX-1 on RoboCasa GR-1 Tabletop and 48.0 percent for the next-highest result in the paper’s comparison table. The 10.7 percentage-point difference is specific to that simulated benchmark. The paper identifies GR00T N1.5 as the 48.0 comparator. AWS’s October 28 partner account
In its benchmark appendix, the report describes 24 GR-1 Tabletop tasks and 50 evaluation episodes per task. The score averages across the defined tasks. The authors also use GR00T results from the official GR00T repository and describe how they reproduced some other baselines. Consequently, the setup and the method of averaging matter when comparing that result with the ALLEX experiment, which uses a physical robot.
AWS also refers to a training compute figure from RLWRLD. The company’s technical blog ties the roughly 20 percent comparison to its own checkpoint trajectory against GR00T N1.5. The comparison addresses training resources for those benchmark results. A warehouse installation has separate hardware, integration and operating costs. The site’s accounting would also show which resources are shared across stations and which belong to this particular step.
A reader can therefore retain the benchmark name, model checkpoint and comparator with the reported number. This is especially useful when later work tests a new release or adopts another definition of success. The first question is whether the two results refer to the same task and conditions.
The AWS account explains how RLWRLD uses cloud compute and storage to train RLDX-1. It describes training on NVIDIA H200 instances, distributed scheduling and shared storage. AWS also reports that RLWRLD became a validated partner and made two offerings available on AWS Marketplace.
These details help explain how a robotics company can process data from multiple sources. For example, it can train on robot demonstrations and generated trajectories and then adapt the model with task-specific data. A training record can identify which inputs were used and which checkpoint was evaluated.
RLWRLD publishes model code and checkpoint links. Its repository also dates the release of RLDX-1.1 code and checkpoints to November 12, 2026. Version information matters because the May paper’s results apply to the setup and model preparation it describes. A repeat experiment can preserve the code revision, weights and protocol together. RLWRLD’s model repository
A customer evaluating warehouse robotics also needs to know how the system would fit its own operation. That can include the required robot, sensors, integration and support. A software distribution channel explains how a customer can obtain an offering. A site report would document how the configured system performed.
On November 9, RLWRLD unveiled the first major release from its DexBench initiative. The company describes 15 tasks across four industry environments, including logistics. It built the benchmark on NVIDIA Isaac Lab-Arena for simulation. Those scenarios give developers common conditions for evaluating robotic manipulation. RLWRLD’s November 9 DexBench announcement
The November materials discuss both changes to tasks and changes to the model. RLWRLD says the initial DexBench baseline results use RLDX-1.1. Separately, its May report presents the earlier ALLEX and GR-1 Tabletop experiments. A dated record makes it possible to follow those results as the evaluation suites develop.
The company also describes plans to compare simulation results with physical robot performance. That is a defined next step. For example, a virtual parcel can be assigned different properties, while a physical trial can measure what happens with the package and tools actually in use.
NVIDIA announced the same day that an investment in RLWRLD was planned. The statement adds information about the companies’ relationship at a point in time. The benchmark’s measured outcomes and an eventual investment transaction each have their own records. NVIDIA’s November 9 investment announcement
CJ Logistics had already reported a robotics deployment on September 3. It said two dual-arm humanoid robots at the Yangji Olive Young logistics center in Yongin were inserting cushioning into boxes. The release lists partner roles across hardware, robot hands and foundation-model work. CJ Logistics’ September 3 deployment announcement
The cushioning example shows a defined station in a customer operation. Staff then place ordered goods in the prepared boxes, according to the release. CJ also identifies RLWRLD as a foundation-model partner within its wider robotics work. Its description of the station centers on cushioning placement.
The later parcel-orientation project poses another warehouse robotics question. It asks whether a system can handle packages so the waybill ends up facing up. The partners’ stated plan is to validate such tasks in an operating environment. A result for that proposed task would have its own model, hardware and protocol.
Other logistics projects can have yet another handling unit. Seoulz’s coverage of Coupang and Contoro describes unloading boxes from vehicles and containers. For a reader comparing these projects, the work unit matters because each station has a different flow of objects and staff. Seoulz’s coverage of Coupang and Contoro
For warehouse robotics, a successful turn is one event in a longer operation. A parcel must reach the next station in the right place and at the right time. Therefore, a trial could define success by the final label orientation and the condition of the package when it moves on.
The evaluation can record how the task changes across packages. For example, the grip needed for a rigid carton may differ from the grip for a flexible package. Overlapping objects can also affect how a robot reaches a target. A useful count identifies which objects and starting positions were included.
A site trial could separately track how often staff intervene, how long recovery takes and what happens to throughput during interruptions. The same observation period can capture successful handoffs, exceptions and the work needed to resume. Then the task result can be read in the context of the station’s operation.
These are illustrative measures for the announced use case. They give each outcome a defined unit, such as successful orientations per attempt or handled parcels per operating hour. Actual results would come from the trial records and their stated conditions.
A warehouse robotics buyer can consider the hand, arm and model alongside the existing conveyor and scanning system. Changing the handling step may also affect the people working before or after it. That is why the scope of a proof of concept matters to the operating team.
Suppose the intended next station is a scanner. A trial might record whether the label is visible when the parcel reaches it. Separately, maintenance staff might need to track how quickly the line resumes after a missed grasp. Both observations can help explain the result for the operation.
Commercial evidence also develops in stages. A partner may fund an evaluation, prepare a proposal for another facility or decide to use a system routinely. Each event can be described with the agreement, site and work it covers. A future proposal can therefore use results from a defined trial to support its own assumptions.
Costs become clearer with that same detail. For example, equipment installation and integration occur at a particular site, while staff training and support can continue over time. A prospective customer can relate those commitments to the measured handling volume and the conditions under which the trial ran.
RLWRLD robot dexterity can now be followed through several public records. The technical report provides methods and scores for defined tasks. AWS describes model training and the partner offering. The September CJ Logistics announcement records a joint development plan and a proposed on-site test.
The conveyor experiment offers a concrete result at the first of those stages. It shows how a task-adapted model performed across four speed settings on ALLEX. The counts and the rule for success are what make that result comparable with a future experiment.
For the announced warehouse work, the next evidence would be a field report that identifies the equipment, task, operating conditions and outcomes. For example, results for successful orientations, interruptions and recovery would answer questions that matter to the station. A later operational decision would add its own scope and date.
The parcel has to reach the scanner in a usable position. A model score can inform the design of that test. Then the site’s own record would show how the handling step fits the work around it.
The bitsensing traffic radar project at Porta Nuova gives a road operator several ways to…
Korea innovation procurement can bring an early product into a public facility where staff have…
Imagine a move-out list with an old refrigerator, a small monitor and a pair of…
Korea technology IPOs can bring a new kind of question into a founder’s laboratory. A…
Korea park reservations become easier to understand when you can name what you want to…
An apartment's heating bill begins well before warm water reaches the floor. Heat may start…
