NVIDIA Builds Bridges Between Korean Chips and More Adaptive Robots
NVIDIA is expanding its presence in Korea around chips, data centers, and agent security. Meanwhile, researchers from MIT and NYU present FLEX, a method to more finely correct robot behavior from natural language instructions.
Two separate announcements published in early November 2026 highlight how AI now plays a role as much in data centers as in robot movements. On one side, NVIDIA deepens its collaboration with South Korea around chips, models, and agent security. On the other, researchers from MIT and New York University present FLEX, a system that adjusts a robot's trajectories based on conversational corrections, with the simple idea of better understanding what the user really wants to avoid, move, or protect.
In Korea, NVIDIA prepares hybrid racks for inference
On November 8, 2026, NVIDIA announced a collaboration with Rebellions and SK Telecom to connect the upcoming Rebellions R200 processors to NVLink Fusion, aiming for deployment in 2028. The project is part of a South Korean technological sovereignty strategy, supported by the Ministry of Science and ICT, and targets systems capable of combining NVIDIA chips and Korean NPU accelerators within the same rack architecture. In short, each component retains its specialty, but all share the same chassis.
The chosen scheme relies on so-called disaggregated inference: NVIDIA GPUs would handle the prefill phase, while Rebellions NPUs would manage decoding. This distribution aims to improve throughput and energy efficiency on very large models, including models with several trillion parameters [to be verified regarding the exact scale of targeted models]. MGX provides, on its side, a common base for rack format, power supply, and cooling, facilitating data center planning as well as industrial integration.
The Korean bet combines hardware, models, and training
SK Telecom does not just host the machines. The operator plans to connect this infrastructure to its proprietary models, including A.X K2, as well as to public services developed under the AI for All program. NVIDIA is also to support three selected consortia and four Korean foundational model builders, with optimization recipes based on Dynamo, an open-source distributed inference framework. The goal is to accelerate the production deployment of Korean open-source models. Again, the promise is not only technical; it concerns the entire chain, from silicon to the final service.
The training aspect is not trivial. NVIDIA announces certification programs, hackathons, and technical seminars to strengthen GPU usage within local teams. This “skills” layer is as important as the servers: without operators capable of running them properly, the beautiful racks remain mostly expensive furniture. Bae Kyung-hoon, South Korea’s Deputy Prime Minister and Minister of Science and ICT, stated that this would allow NPUs designed in Korea to “enter the computing infrastructure market alongside NVIDIA.”
FLEX aims to correct robots like rephrasing an instruction
At MIT and NYU, FLEX, standing for Flexible Feature Editing Pipeline, tackles a more subtle but very concrete problem: how to make a robot understand what a human wants to correct without imposing an endless list of rules. The system uses natural language corrections to modify the “features” a robot monitors in its environment. In the example given by the researchers, telling a robot carrying coffee not to spill it on a laptop does not just lead to avoiding that laptop; FLEX also infers a broader preference to avoid nearby electronic devices, such as a phone placed right next to it.
The pipeline separates preference formulation and behavior optimization. A correction goes through three steps: ADD checks if existing features are sufficient; if not, the language model proposes a missing variable and generates the associated code. GENERALIZE replaces an overly specific rule with a broader intention. MASK removes elements that have become irrelevant. In the bowling ball example cited in the paper, a constraint limited to a cup becomes a more general avoidance instruction for fragile objects. This sequence aims to limit half-true, overly precise, or plainly misplaced representations.
Promising results, but still mostly in simulation
The researchers say they tested FLEX on 9 x 9 navigation grids and on simulated tasks with a seven-degree-of-freedom Franka arm, around a coffee cup, a lit candle, and a bowling ball. In the seven navigation scenarios, FLEX achieved the highest average normalized reward score. In the case requiring all three operations, it reached 0.99 versus 0.69 for ALGAE, 0.83 for MFE, and 0.78 for the unstructured baseline. Regarding the user study, 12 participants provided textual corrections over three blocks of five sequences; FLEX received the highest average rating across five evaluated dimensions, with a score of 6.3 out of 7 for intention understanding.
One significant limitation remains: the human study was not conducted in a closed loop on a real robot. Participants observed simulated trajectories because some comparison methods produced erratic movements. The authors also acknowledge that FLEX remains limited to quasi-static manipulation of rigid objects, whose relevant properties can be described in language. Dynamic tasks, deformable objects, or highly personal preferences will require something else. For a domestic robot, this means conversation helps but does not yet replace practical common sense.
Key points
- NVIDIA, Rebellions, and SK Telecom aim for deployment in 2028.
- FLEX was presented by MIT and NYU on November 9, 2026.
- FLEX score reaches 0.99 in a full navigation scenario.
- 12 participants evaluated the system in the user study.
- SK Telecom connects systems to A.X K2 and AI for All.
- The FLEX project remains limited to quasi-static rigid objects.
By the numbers
- 2028 — target year for deployment of Rebellions/NVIDIA systems in Korea.
- 0.99 — average normalized reward of FLEX in a three-operation scenario.
- 12 participants — panel of the user study on FLEX.
- 5 seconds — announced duration for the complete ADD-GENERALIZE-MASK loop.
- 9 x 9 — size of navigation grids tested by the researchers.