Whole-Body WAM - Unitree G1 loco-manipulation
Extending FastWAM from tabletop manipulation to 72-D whole-body physical action prediction on Unitree G1 + Wuji hands.将 FastWAM 从桌面机械臂任务扩展到 Unitree G1 + Wuji 的 72 维全身物理动作预测。
Projects
项目
π₀.₅ on Franka - flexible manipulation
An end-to-end OpenPI adaptation for Franka: DROID-style demonstrations, LeRobot v2 data, 8×A100 full fine-tuning, WebSocket policy serving, and real-robot execution.面向 Franka 的 OpenPI 全链路适配：DROID 风格示范、LeRobot v2 数据、8×A100 全量微调、WebSocket 策略服务与真机执行。
Audio-L3A - efficient audio class-incremental learning
A frozen-backbone analytic learner reaching 45.16% mAP on five-phase AudioSet-50, 2.65 points above LwF at roughly one-ninth of offline training time.冻结 backbone 的解析式持续学习方法：AudioSet-50 五阶段达到 45.16% mAP，领先 LwF 2.65 个百分点，总训练时间约为离线上界的 1/9。
NanoMemAgent — modular agent framework with layered memory
A modular agent runtime with a ReAct loop, plug-in skills, MCP tool ecosystem, and an L1–L5 memory stack (vector + BM25 + knowledge graph).模块化 Agent 执行框架：ReAct 推理循环、插件化 Skill、MCP 工具生态，以及 L1–L5 分层记忆栈（向量 + BM25 + 知识图谱）。
PocketLLM — 0.2B LLM trained from scratch, full pipeline
A 0.2B-parameter MoE decoder-only model built end-to-end on a single GPU: pretraining + SFT + DPO / PPO / GRPO alignment. Benchmarked on C-Eval and CMMLU.单卡从零训练的 0.2B 参数 MoE Decoder-only 模型：预训练 + SFT + DPO / PPO / GRPO 对齐全链路。在 C-Eval、CMMLU 上评测。
learn_embodied_papers — reading notes in the open
An open notebook of embodied-AI paper-reading notes: VLAs, manipulation policies, diffusion and flow-matching methods.公开维护的具身智能论文阅读笔记：VLA、操作策略、diffusion 与 flow-matching 方法。
Titan — capstone with PredicTx Health
Management control plane for Runbeam Harmony proxies — medical-data interoperability agents built with PredicTx Health, a precision-oncology company. Capstone team of ten.为 PredicTx Health（精准肿瘤学公司）的 Runbeam Harmony 代理服务构建管理控制面——医疗数据互操作代理的 Capstone 项目，团队十人。