Projects

Tools

LLM 训练显存估算器

拖动层数、hidden size、序列长度、TP/DP/ZeRO,粗略估算单卡训练显存由哪几部分组成、能不能放得下。

Interactive · LLM Training

Research & Course Projects

QDGS: 3D Scene Reconstruction with High-Dimensional Surface Primitives

Extended 2D Gaussian Splatting by replacing planar primitives with high-dimensional curved surface primitives. Designed geometric routines for ray-surface intersection, differentiation, and bounding box computation. Improved reconstruction error from 0.78 to 0.72 while using fewer primitives.

Dec 2024 · 3D Reconstruction · Gaussian Splatting

CyberDog 2 Dribbling with Deep Reinforcement Learning

Trained quadrupedal and bipedal dribbling policies for CyberDog 2 with reinforcement learning. Implemented a vision-based goalkeeper module and a complete 2v2 match environment; placed 2nd of 16 teams in a university robotics competition.

Jun 2024 · Reinforcement Learning · Robotics

LLM-Enhanced RPG Game with Context-Aware NPC Agents

An LLM-driven RPG prototype built with GPT-4o and RPG Maker. Prompt templates and game-state interfaces let NPCs condition their dialogue on player progress, dialogue history, and task context.

Jun 2024 · LLM Agents · Game AI