Bristol-VLABristol-VLA 项目
Fine-tuning π0.5-DROID with tactile images用触觉图像微调 π0.5-DROID
I lead the VLA workstream, covering demonstration data, policy fine-tuning and deployment interfaces. I work with colleagues on the hardware. For the tactile runs, we combine two tactile RGB images into a single third input for π0.5-DROID, alongside the front and wrist camera views.我牵头做 VLA 这部分，负责示教数据、策略微调和部署接口，硬件则和同事一起做。用到触觉的训练里，我们把两路触觉 RGB 图像合成一张，作为 π0.5-DROID 的第三路图像输入，和前视、腕部相机的画面一起送进模型。
- My role我的角色
- Project Lead · VLA Development项目负责人 · VLA 研发
- Research setting研究环境
- Bristol Robotics Laboratory
- Status进展
- 3 fine-tunes complete; robot comparison pending三组微调已完成，真机对比待做
- 01
Prepare demonstrations整理多模态示教
30 demonstrations · 8,116 selected frames30 段示教 · 筛选后 8,116 帧
Implementation展开技术细节
Four RGB streams capture front, wrist and two tactile views alongside robot state and gripper intent. Rule-based checks retain 253 continuous segments for the visual and tactile baseline experiments.四路 RGB 记录前视、腕部和双路触觉，同时保存机器人状态与夹爪意图。通过规则校验，为视觉基线及触觉版本保留 253 个连续片段。
Franka / ZED 2i / D435 / tactile RGB
- 02
Adapt the VLA policy微调 VLA 策略
3 independent runs · 5,000 steps each3 组独立实验 · 各 5,000 步
Implementation展开技术细节
Native π0.5-DROID is adapted with LoRA. Two tactile RGB images share a third policy image slot; training and inference use the same adapter. The three runs cover visual, visual–tactile and reviewed-export data.基于原生 π0.5-DROID 进行 LoRA 微调。双路触觉 RGB 合成为策略的第三路图像，训练与推理共用适配器；三组实验分别采用视觉、视觉触觉和人工审核导出数据。
OpenPI / LoRA / H15 / 8× H100 access
- 03
Prepare deployment衔接推理与部署
RTX 5090 · consistent policy inputsRTX 5090 · 一致的策略输入
Implementation展开技术细节
An RTX 5090 hosts the documented OpenPI inference setup. Each checkpoint must be deployed with its matching image composition and normalisation. The robot platform uses native libfranka teleoperation and a DROID impedance-control interface.在 RTX 5090 上搭建 OpenPI 推理环境，部署时配套使用 checkpoint 对应的图像合成方式和归一化文件。机器人平台使用原生 libfranka 遥操及 DROID 阻抗控制接口。
OpenPI / libfranka / checkpoint + adapter
- 04
Design the robot comparison准备真机对比评测
Shared red-cube protocol · in preparation统一红方块任务 · 正在准备
Implementation展开技术细节
The planned comparison fixes the setup, reset and success criteria, then records every trial and intervention. Data checks and completed training are documented; a comparable closed-loop trial set is still outstanding.计划固定实验环境、复位条件和成功判据，逐次记录试验及人工干预。数据检查与训练完成已有记录，可比较的真机闭环试验集仍待完成。
Fixed conditions / trial records / robot evaluation
The dataset contains 30 demonstrations.数据集包含 30 段示教。
Read the full VLA case study查看完整 VLA 项目案例