计算机视觉 · 机器人 · 大模型后训练 — 竞赛级方案 · 代码级复盘 · 开源。
🎯 currently: edge deployment of detection / fusion models（检测 / 融合算法的边缘部署）
| Competition | Approach | Result |
|---|---|---|
| Qwen-Cup-Multimodal-RL | Two-stage post-training of Qwen3.5-0.8B (multimodal): teacher-distillation SFT from Qwen3.6-27B (113K distillate → 24K curated) + small-dose GRPO RL (+2.0 over the SFT baseline); verbatim reward–eval alignment, statistical submission budgeting | Preliminary 20/337 → Finals leaderboard 9/20, climbed to 3rd overall after the defense — Qwen Cup Multimodal Reasoning (Alibaba Qwen) · ¥20,000 prize |
| AIC-4D-Radar-Detection | 3D detection from 4D mmWave radar point clouds (radar-only): SFA3D network kept as-is, original 8D→4D SNR mapping + cross-class NMS · 110 FPS · INT8 12.3 MB | National 2nd Prize, AIC 2025 Global Campus AI Algorithm Elite Competition · 75 mAP@0.5 |
| Scientific-Image-Forgery-Detection | Copy-move forgery segmentation in scientific images: frozen DINOv2 + lightweight CNN decoder, conservative dual gating against false positives, honest retrospective incl. gold-solution paradigm analysis | 🥈 Kaggle Silver Medal (1,564 teams, top 5%) |
| Go2_RL | Unitree Go2 quadruped RL: autonomous navigation + locomotion (PPO + reward engineering aligned with the scoring formula + curriculum learning) | Regional Second Prize · National Finals Third Prize (2026 Tencent AI Arena Global Open) |
Every competition gets a code-level retrospective, open sourced — the methodology is worth as much as the result.
Outstanding undergraduate thesis award for an agent project.
📧 webcainiao@gmail.com · Kaggle @web3cainiao
中文
计算机视觉 · 机器人 · 大模型后训练。每场竞赛都做代码级复盘并开源——复盘方法论和名次一样值钱。
- 千问杯多模态推理 — Qwen-Cup-Multimodal-RL：Qwen3.5-0.8B 两阶段后训练——教师蒸馏 SFT（Qwen3.6-27B 教师，113K 蒸馏 → 24K 精选）+ GRPO 小剂量强化（同口径较 SFT 基线 +2.0；奖励-评测口径逐字对齐、提交额度统计化决策）。初赛 20/337 晋级 · 决赛榜 9/20 · 答辩计入总评逆风翻盘至最终第 3 名 · 2 万元奖金（阿里巴巴千问杯）。
- 4D 毫米波雷达 3D 检测 — AIC-4D-Radar-Detection：基于 SFA3D（网络沿用原版），原创 8D→4D SNR 映射 + 跨类别 NMS，110 FPS，INT8 量化后 12.3 MB。AIC 2025 全球校园算法精英大赛全国二等奖 · 75 mAP@0.5。
- Kaggle 银牌 — 科学图像伪造检测（1,564 队，前 5%）：冻结 DINOv2 + 轻量 CNN 解码器做 copy-move 伪造分割，保守双门控压误检，附诚实复盘（做对/做错 + 金牌方案范式分析）。
- 四足机器人强化学习 — Go2_RL：Unitree Go2 自主导航 + 运动控制（PPO + 奖励工程深度对齐评分公式 + 课程学习）。腾讯开悟全球公开赛区域二等奖 · 全国三等奖。
currently：检测 / 融合算法的边缘部署。
其他：智能体课题获优秀学士毕业论文。