Platform, PM and Optimization now sit beside the AI Engineer

CMES Robotics currently lists separate openings for Physical AI Engineer, Physical AI Platform Development, Physical AI Planning/PM and Physical AI Combinatorial Optimization. The four public pages show September 30 deadlines.

The Engineer role designs VLA models, Robot Action Policies and imitation-learning systems and validates them in simulation and on real robots.

The other three roles make the structure more interesting: the company is not only adding model researchers, but separating the capabilities needed to apply those models repeatedly in industrial environments.

The Platform role connects Teleoperation to Inference in one loop

The platform opening explicitly covers data acquisition through Teleoperation and Hand Retargeting, data management, model training, validation, deployment and real-time inference across heterogeneous robots.

It also covers synchronized RGB-D, robot-state and end-effector data, Dataset/Data Format/Model Weight Versioning and cross-platform interfaces for Single Arm, Dual Arm and Humanoid systems.

The purpose is not simply another model. It is a development environment in which Physical AI engineers can repeatedly collect data, train and deploy across different robots and tasks.

The Physical AI Center PM decides which tasks and robots enter the system first

The Planning/PM opening explicitly includes strategy for building and operating an internal Physical AI Center. It covers Cell/Zone plans for Single Arm, Dual Arm and Humanoid robots and phased introduction roadmaps.

The role also identifies automation tasks in industrial sites, prioritizes them, and plans a Real-to-Sim-to-Real process for data acquisition, model development, validation and field deployment. Teleoperation, Motion Capture, simulators, end effectors and AI training infrastructure are treated as Center components.

A public job posting cannot establish the Center's actual scale, budget or launch date. But it does show a role designed to coordinate teams, robots, datasets and customer tasks as a system rather than a one-off research project.

Optimization connects model capability to real customer constraints

The combinatorial-optimization opening covers palletizing and depalletizing plans, robot work sequencing and task allocation for real industrial deployments.

A robot being able to pick an object is not enough to make a production system. The system must decide sequence, placement, workload allocation and how to satisfy cycle-time and space constraints.

That creates a separate talent layer between model performance and operational performance: people who translate robot capability into throughput and customer KPIs.

After a KRW 6.91B field contract, repeatability becomes a different problem

In May 2026, CMES Robotics disclosed a KRW 6.91 billion contract with Coupang Fulfillment Services to install a logistics robot automation solution. The amount was 52.93% of recent annual revenue, with a contract period from July 1, 2026 through March 2, 2027.

Delivering one large customer project and repeating the same Physical AI capability across multiple customers and sites are different operating problems.

As deployments multiply, rebuilding data collection, model management and deployment for each robot becomes expensive. Shared platforms, Center-level processes, reusable task definitions and field optimization become more valuable.

The talent competition is no longer only about hiring the best AI researchers

VLA, imitation learning and robot policy research remain important. But as deployments expand, a labor structure forms around the model.

One group builds robot intelligence, another owns the data/model lifecycle as a platform, PM defines task and robot priorities through the Center, and optimization engineers translate customer constraints into executable systems.

Counting AI researchers alone can therefore miss a large part of the capability needed for Physical AI commercialization.

BANSEOG VIEW | Physical AI is moving from Model to Operating System

The common signal across the four CMES openings is that the structure for repeated application is becoming more specialized than the model itself.

Engineer creates intelligence. Platform standardizes the data-to-deployment loop. Center/PM selects tasks and rollout order. Optimization brings customer constraints and KPIs into the system.

The next Physical AI competition may be not only who can build the best demo, but who can build an operating system that repeats learning, validation and deployment across many tasks and customer sites.

BANSEOG VIEW

Banseog View — Model → Platform → Operating System

CMES Robotics is recruiting a Physical AI Engineer alongside separate Platform, Center Planning/PM and Combinatorial Optimization roles.

The Platform role links Teleoperation and Retargeting with data management, training, validation, deployment and inference in a repeatable environment.

As field contracts grow, Physical AI talent strategy may expand from researcher headcount toward ownership of an operating system that works across robots and tasks.

SOURCES

Primary sources and references

Public openings do not reveal CMES Robotics' full organization, hiring volume, or the actual build progress, budget or launch timing of the Physical AI Center. This article therefore does not claim the Center is already complete or that these four roles represent the company's entire Physical AI workforce. Banseog has previously covered teleoperation and data loops; the additional value here is the company-level separation of Engineer→Platform→Center/PM→Optimization as a commercialization organization. ‘Model → Platform → Operating System’ is Banseog analysis.