AI in the Trenches: Clear Signals from Science Alive 2025
At the IBS Science and Culture Center in Daejeon on the 11th, researchers sat down for the "AI Era Science Scene" Deep Talk during the "Hot & Deep" section of Science Alive 2025. The session cut through hype and focused on what is actually working, what isn't, and what needs to change.
Noh Ju-won, Head of External Affairs at KIST, moderated. Panelists included Won Hong-in (KITECH), Professor Park Hae-won (KAIST), Yang Kyung-wook (KRICT), Heo Mu-yeong (IBS), and Professor Kim Won-hwa (POSTECH).
Humanoids: Hope vs. Proof
Questions about humanoids came fast. Professor Park didn't sugarcoat it: despite headlines about deployments at factories and fulfillment centers, there's little verified evidence on performance or real substitution of human labor.
Startups need investor attention, so messaging leans optimistic. The core issue isn't ambition-it's data.
The Data Problem in Robotics
For large language models and image models, the internet is a massive training set. Robotics doesn't have that advantage. YouTube shows people doing tasks, not robots executing them under constraints.
According to Park, high-quality robot data requires people teleoperating robots so the system can learn by following. That collection loop is slow and expensive compared with labeling images or text. It's the bottleneck, not just model architecture.
From Lab to Factory: Accuracy and Cost
Even with better models, industrial floors demand brutal reliability. "To use manipulators on site, operational accuracy needs to be at least 99.99%," Park said. Anything less and operators reject the tool.
Won Hong-in agreed: robot systems are still too costly to beat labor in most cases. He expects a few large manufacturers to prove a repeatable model first. Only then can suppliers and integrators retool the ecosystem-which takes time.
AI in Research Workflows: Gains and Gaps
Professor Kim noted a real productivity boost in academic writing. Drafts move faster, and language barriers weigh less. But AI still struggles with broken or irregular structures-repair tasks with messy geometries still call for human judgment.
Yang shared a practical win: AI can translate papers into Korean and reshape them into press releases with the right framing. Still, it may "interpret" figures and invent maximum values. Review is non-negotiable.
Practical Guidance: Reduce Hallucinations
Heo Mu-yeong's advice was blunt: models guess when they don't know. If you don't load the right manuals or references first, you'll get confident nonsense. Ground the system with domain material, then ask questions.
This is retrieval before generation. Treat it like instrument calibration-no context, no trust.
Actionable Takeaways for Researchers
- Separate claims from proof. Ask for metrics, failure rates, and deployment duration on any robot "success."
- Expect data collection to dominate robotics timelines. Budget for teleoperation, evaluation, and iteration.
- Aim for 99.99% where safety and uptime matter. A demo isn't a deployment.
- Use AI for drafting, translation, and structure. Keep humans on ambiguous edge cases and figure interpretation.
- Prevent hallucinations: load manuals, SOPs, and datasets first; then query. Log sources for verification.
Why This Matters Now
The infrastructure for robot learning is catching up, but it isn't there yet. Meanwhile, AI is already useful in research pipelines-writing, summarizing, and planning-if you keep it grounded and supervised.
As Noh closed, "The AI era is an undeniable, major trend. Let's be the ones who control AI, not the ones controlled by it."
If you want structured practice in prompts, retrieval, and verification for research work, here's a useful starting point: AI courses by job.
Related reading on training gaps: Programming by Demonstration (robot learning).
Your membership also unlocks:
