Multimodal learning for physical robots面向真机的多模态学习

How can contact feedback help a VLA policy handle slippery, deformable and contact-sensitive objects?触觉反馈如何帮助 VLA 策略操作易滑、可变形及对接触敏感的物体？

I lead Bristol-VLA and its VLA workstream, using dual-Franka leader–follower demonstrations to adapt π0.5-DROID with paired tactile RGB. Three independent LoRA fine-tuning runs are complete; data consistency checks and episode-level holdouts prepare the work for a shared closed-loop robot comparison.我负责 Bristol-VLA 项目及 VLA 路线，基于双 Franka 主从示教，将双路触觉 RGB 接入 π0.5-DROID。三组独立 LoRA 微调已完成，通过数据一致性检查和 episode 级验证集划分，为统一的真机闭环对比做准备。

Can event-based touch detect surface defects in real time on a manufacturing line?事件驱动触觉能不能在制造流程中实时发现复合材料表面缺陷？

I led the seed study and the successful £150k follow-on proposal. Current work combines tactile hardware development with ABB RobotStudio planning, surface-normal tracking and constant-force scanning on ABB robots.我以 PI 身份完成种子研究，并主导后续 15 万英镑项目申请并获批。目前的工作将触觉硬件研发与 ABB RobotStudio 规划、曲面法向跟踪及 ABB 真机恒力扫描结合起来。

Do people change their preferred touch speed according to how a texture feels?面对不同的纹理感受，人会不会主动改变自己偏好的触摸速度？

I led a human study across natural textures, velocities and contact forces, then modelled how perceptual dimensions predict preferred exploration speed.我设计并完成了自然纹理、不同速度与接触力下的人体实验，再用感知维度预测人偏好的探索速度。

Which interaction pattern gives a robot the clearest evidence about texture?滑动、旋转或组合动作，哪一种能让机器人获得更清楚的纹理信息？

As co-first author, I compared six individual and compound movements using NeuroTac. This research also spans SCNN development and deployment on SynSense Speck2f, connecting event-stream processing with on-chip texture inference.我以共同一作身份使用 NeuroTac 比较了六种单独或组合动作。这条研究线还包括 SCNN 开发及 SynSense Speck2f 部署，将事件流处理与片上纹理推理连接起来。

Can a robot infer the perceptual dimensions of a texture it has never encountered?面对训练中从未见过的纹理，机器人能不能预测人会从哪些维度感知它？

As first and corresponding author, I built a dataset linking 24,000 NeuroTac interactions across 50 textures with nine human perceptual attributes, then evaluated a CNN–TCN on unseen textures.我以第一作者、通讯作者完成这项研究，将 50 种纹理的 24,000 次 NeuroTac 交互与人的九项感知属性对应，并用 CNN–TCN 评估对未见纹理的预测。