Embodied-AI & Human-Robot Interaction Research · Remote 具身智能与人机交互研究 · 远程弹性工作制
- RynnBrain VLM × OmniGibson evaluation pipeline: Removed simulator privileges one by one — ground-truth segmentation → GSAM open-vocabulary perception, ground-truth pick height → depth, ground-truth container pose → perception — with CuRobo motion planning, reaching 80% BDDL success on 20 seeds; then showed that a teleport-based stabilizer was inflating BDDL success and rebuilt the success judge (strict prompt, median-of-3). RynnBrain VLM × OmniGibson 评测管线：剥离仿真特权信息（真值分割 → GSAM 开放词汇感知、真值抓取高度 → 深度、真值容器位姿 → 感知），配合 CuRobo 运动规划，在 20 个种子上达到 80% BDDL 成功率；随后证明瞬移式稳定器虚高了 BDDL 成功率，并重建了成功判定（严格提示词、三次取中位数）。
- Humanoid voice interaction: Co-deployed a real-time domestic voice stack on a humanoid robot for water-fetch and reception scenarios using Alibaba Cloud TTS/STT, Agora RTC, and TEN Framework, researching 3C noise suppression and microphone directivity. 人形机器人实时语音交互：在人形机器人上联合部署迎宾与递水场景的全栈国产化实时语音系统（阿里云语音合成/识别、声网 RTC、TEN Framework），研究软硬件边界的 3C 降噪与麦克风指向性。
- Multi-cloud GPU infrastructure: Collaborated with several partner infrastructure teams on GPU container scheduling, heterogeneous domestic/NVIDIA cluster orchestration, and vLLM inference pipelines. 多云异构算力协同：与多家合作方基础设施团队协同排查 GPU 容器调度、国产芯片与 NVIDIA 混合集群编排及共享算力推理管线。