Work

Llm의 행동 데이터 생성과 반복 가중치 개선을 통한 로봇 모델 제작

Year: 2026
Type: preprint
Abstract: Building a robot model requires repeated data collection, policy training, failure analysis, and modification. This work places the LLM in the role of model producer and proposes a production loop connecting behavior-data generation to improvement of the trained model's weights. A person initially specifies the desired model and provides feedback from real-world use of the resulting model, while AI performs the intermediate generation, training, modification, and testing. In the first case, RaccoonBot, demonstrations from a tool-using teacher train a geometrically integrated GRU policy, and the LLM designs, implements, and executes correction procedures to produce candidate weights. Numerical updates use AdamW, demonstration-frequency-based distance estimation, and Newton-based CLOSE correction. In stored simulated development evaluations, successes for the selected seed increased from 238/252 to 252/252 unique starts and from 12/24 to 24/24 conditions for seven- to eight-step programs, comparing the existing approach-distance policy with the combined candidate. This limited-task case demonstrates an implementation path, not fully autonomous model production or independent generalization. Follow-up research will compare improvement mechanisms and human intervention under matched budgets and investigate reuse of the production loop across robots and purpose-specific models. Version 0.8.3: bilingual research preprint, Korean (16 pages) and English (13 pages). The introduction and Section 6.3 integrate a long-term hypothesis: improved LLM physical perception and reasoning may support analysis of parameter-behavior relationships, targeted direct weight modification, and purpose-specific model production. Direct modification alone is defined at the improvement stage, not as eliminating initial demonstrations/training, tests, or real-world feedback. Section 6.3 connects the hypothesis to predictive behavioral checks, matched-budget controls, retention of existing capabilities, experience reuse, and robot-specific adaptation. The existing production loop and distinction between current numerical fitting and future direct editing are retained. Methods, results, all tables, conclusion, references, and evidence appendices are unchanged. No new experiments were performed. Previous version 0.8.2 remains available at https://doi.org/10.5281/zenodo.24217129. Sole author: 유용현. Research repository: https://github.com/dreaminglizard/llm-driven-robot-policy-learning. CC BY 4.0 applies to the papers in this record, not automatically to linked code, datasets, or model weights. SHA256 paper_ko.pdf: ef0849c4185b6cbc3872a839e905d9095170050518dc78fe2634a72cc644ef19. SHA256 paper_en.pdf: e472279cd223f60a6cca47cf4fca15cce1181b94e30ab5bae03deeecbcf86960.
Author: 유, 용현
Language: English
Open Access status: green
OpenAlex ID: W7229409473
Cites:
Cited by:
Related to: 10