2026-10-4 · Build Model
学会如何搭建一个基础的模型
Natural-language instruction
Environment & plan checks
Execute · Reject · Clarify
MSc Dissertation · 2026
Environment-Aware Feasibility Planning for Local LLMs in MuJoCo
A MuJoCo tabletop benchmark tests whether local LLMs distinguish executable, infeasible and ambiguous instructions when the language stays fixed but the environment changes.方法、贡献与边界
168tasks
84matched pairs
9,828experiment records
My contribution
I designed seven controlled failure categories and implemented a deterministic feasibility verifier. Structured feedback supports up to three rounds of plan repair and blocks plans that remain infeasible from execution.
Evaluation
Across six local models and 126 held-out tasks, verified planning reached 95.2–100% matched-pair success for five models; the full configuration recorded 0% simulator crashes. This is planning-feasibility evaluation, not robot policy training.
