Robotics is shifting from hardware sales to data-driven service models, while Korea's open-weight AI models (LG AI Research, SKT, Upstage) differentiate on licensing and privacy against Chinese rivals. The case gives overseas companies and buyers evaluating physical AI or model partnerships concrete evidence of technical maturity and trustworthy partner candidates.
Physical AI Moves from Demos to the Factory Floor
Robots are moving beyond repetitive tasks into complex on-site work, and the business model is shifting with them —from selling hardware to selling data-backed services
Key Issues
- Whole-body control and multi-robot teamwork: Google DeepMind released Gemini Robotics 2 on 30 July, a three-model family that drives legs, torso, arms and fingers under one learned policy and lets different robots split a workflow. It ships with ASIMOV-Agentic, a benchmark that measures whether a robot refuses unsafe commands and asks for human help.
- Planning while moving cuts cycle time: MIT's VLASH framework lets a robot plan its next motion while it is still executing the current one, making tasks two to three times faster with only a small trade off in precision.
- From selling robots to selling a service: Maum AI says customers cannot handle data collection, GPU training and retuning themselves, so from next year it will supply robots and AI as a service. Project prices have tripled or quadrupled as inquiries shift to manufacturing, construction and defense.
Korea's Open-Weight Models Head into a Second Evaluation
Ahead of the second evaluation of the government's Independent AI Foundation Model project, Korean open-weight models lead on tasks but lag in inference efficiency
Key Issues
- A 750-billion-parameter model under an open license: LG AI Research released K-EXAONE 2.0 on Hugging Face, a Mixture-of-Experts model that activates only 37 billion of its 750 billion parameters per query and carries the commercially usable Apache 2.0 license. It averaged 70.1 across 24 benchmarks, up from 63.3.
- Strong in Korean, behind on efficiency: Across seven shared benchmarks, SK Telecom's A.X K2 led four and Upstage's Solar Open 2 led three, while K-EXAONE 2.0 led none —though the tests center on Korean language and culture rather than global comparison.
- Chinese rivals gain on cost but not on privacy: Kimi K3 activates 16 of 896 experts, putting active parameters at about 3.7% of the total, against 4.1–6.0% for the Korean models. Neither Kimi K3 nor GLM-5.2 offers an in-app opt-out from training —users must email a request —and Kimi's policy allows clipboard collection.
For further details, please refer to the source links provided in the file.