Robbyant 发布 LingBot-Touch 触觉基础模型，跨传感器迁移误差降低 97%
做机器人触觉的朋友看这个：Robbyant 把 4983 小时触觉数据训成一个冻结可迁移的基础模型，在新传感器上误差从 5.89mm 降到 0.16mm，还带动 VLA 成功率从 57% 升到 75%。权重代码全开源，实验室可以直接上手验证。
Robbyant 团队发布机器人触觉基础模型 LingBot-Touch，预训练数据达 4,983 小时真实触觉数据，解决机器人手上不同触觉传感器之间模型难以迁移的问题。模型在从未见过的 XHand1 机器人手上测试，主模型全程冻结，只训练一个小型附加模块来定位指尖按压位置。
迁移能力的关键在于将所有传感器统一映射到同一个虚拟指尖表征，训练时还随机改变触点密度和位置，防止模型记住某一特定传感器。在 XHand1 上，LingBot-Touch 的平均接触中心误差为 0.16 mm，而使用 XHand1 专用适配器的 FTP-1 误差为 5.89 mm，错误率降低约 97%。
更好的触觉直接提升任务表现：在六项真实世界 XHand1 任务中，LingBot-VLA 在无触觉输入和原始触觉输入下成功率均为 57.21%，接入 LingBot-Touch 特征后升至 75.04%；转盘子任务从 0.5% 提升到 97.4%。论文、代码和模型权重均已公开，覆盖 arXiv、GitHub 和 Hugging Face。
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: arxiv.org/abs/2611.12806 Code: github.com/Robbyant/lingb… Weights: huggingface.co/collections/ro…📊 3 ⚡
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include LingBot-Touch, 触觉, 机器人, 开源模型.
Source attribution: curated from elvis (@omarsar0).
