Seoul National University robot AI pictures the result before it moves
A team led by Professor Jaeyoung Do at Seoul National University built Dynin-Robotics, a single model that reads instructions, looks through a camera and then imagines what the scene should look like after it acts. The robot uses that predicted image to generate or pick its next move, a bit like a person picturing where a cup will land before reaching for it. Across four real-world tasks, including moving fruit and stacking colored blocks, it averaged a 78.4 percent success rate, the highest among the NVIDIA and Physical Intelligence models in the comparison. The paper was accepted to NeurIPS 2026, and the team is looking at manufacturing, logistics and medical robots next.
Read the full story at Seoul National University College of Engineering →