Robbyant 的論文提出觸覺基礎模型 LingBot-Touch,並在訓練時未見過的 XHand1 機器手上測試跨感測器遷移能力。
- - 訓練與測試:模型以 4,983 小時真實觸覺資料預訓練。
- 測試時模型保持凍結,只訓練一個小型附加模組來定位指尖受壓位置。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
Impressive paper from Robbyant on physical AI.
(bookmark it)
Robot hands use a range of tactile sensors, and a model trained on one sensor can struggle to transfer to another.
That's a big challenge in robotics right now.
LingBot-Touch is a foundation model to address this challenge. It was pretrained on 4,983 hours of real touch data.
They tested it on the XHand1, a robot hand whose sensor the model never saw. The model stayed frozen. Only a small add-on was trained to find where the fingertip is pressed.
On XHand1, the average contact-center error was 0.16 mm with LingBot-Touch versus 5.89 mm with FTP-1 using an XHand1-specific adapter. That's about a 97% reduction in error.
It transfers because every sensor maps to the same virtual fingertip. Training also varies touch-point density and placement, so the model can't memorize one sensor.
Better touch also makes robots better at tasks. Across six real-world XHand1 tasks, LingBot-VLA averaged 57.21% success both without tactile input and with raw tactile input, rising to 75.04% with LingBot-Touch features.
In simulated RL, it scored highest on all twelve tasks. Plate spinning went from 0.5% to 97.4%.
Has anyone seen a frozen touch model transfer to a new sensor like this?
Paper: https://t.co/NFcvmcV0IS
Code: https://t.co/4uYmMlv3BW
Weights: https://t.co/OUwaODF0Tu