1. Companion Robot
Restful — ROS2 — Gazebo, Unity simulators (private products purchased — model and ROS provided)
: Implementing autonomous driving via the A* algorithm (with provided maps and pathfinding) and randomized decision-making
#QA
Order matters: build the control configuration first, then think about tools like ROS, Isaac Sim, CoppeliaSim, etc. Start by thinking about core code, including simulation methods such as occupancy grid mapping.
If sensing is critical, go with Arduino; if computation is critical, go with Raspberry Pi.
2. Vision
Change brought about by one person's drive (ImageNet)
Going forward, AI will play a role like electricity.
Visual image matching, 3D reconstruction research
China, the lore of the ring — proof through that kind of life rather than mere acquisition
3. AI Implementation of Characters from a Book
Content consumption driven not by a person's reflection or choice, but by content's one-way push = an era of missing humanities
People speak more often than expected without thinking, and in ways completely unrelated to being logical.
A competitor? Naver Webtoon's character AI, released around the same time, focuses simply on giving accurate answers about the (learned) story it knows. In contrast, the book-character AI is being researched with a focus not just on matching the same person's tone but on knowledge and attitudes from that era (point in time), centered on 'what makes Hong Gil-dong feel like Hong Gil-dong.'
To improve the unconditionally positive tone of current LLM responses, fine-tuning on negative-response data is being planned.
(Should have asked why they thought of fine-tuning rather than RAG or prompts...)
# QnA
Fine-tuning for the unconditionally positive bias
The QA list is generated by an LLM, with the human evaluating the accuracy of A.
: Construct a guide to minimize the checker's bias (e.g., structuring relationships among characters)
: Also considering fine-tuning a small LLM to operate in MoE fashion
Unlike distinguishing objects,
build a guide on how to distinguish people
: Even if their appearance is the same, listing how it feels (in dialogue) when oo is missing
4. RL-Based Quantum Coin Trading
RL and trading share structural similarities such as their sequential nature and decision-making within max/min ranges.
5. Open-source CDC Solution Debezium
: composed of the magic number of 5 + alpha
: To build a foundation for collaboration (shared interests, background knowledge), they hold joint study sessions and attend conferences.
6. DR4R Imitation Learning
: implementation focus (.. moveIt)
The Mobile Aloha open source exists, but it requires the same hardware spec, which costs about 50 million KRW and has a steep learning curve.
With Stanford's UMI open source, you only need to develop the gripper hardware (since you only need to know the endpoint), and you can immediately deploy it on a 6-axis robot.
