Tactile perception is a critical component of solving real-world manipulation tasks, but tactile sensors for manipulation have barriers to use such as fragility and cost. In this work, we engage a robust, low-cost tactile sensor, BeadSight, as an alternative to precise pre-calibrated sensors for a pretraining approach to manipulation. We show that tactile pretraining, even with a low-fidelity sensor as BeadSight, can improve an imitation learning agent's performance on complex manipulation tasks. We demonstrate this method against a baseline USB cable plugging task, previously achieved with a much higher precision GelSight sensor as the tactile input to pretraining. Our best BeadSight pretrained visuo-tactile agent completed the task with 70\% accuracy compared to 85\% for the best GelSight pretrained visuo-tactile agent, with vision-only inference for both.

翻译：触觉感知是解决真实世界操作任务的关键组成部分，但用于操作的触觉传感器存在易损性和成本高等使用障碍。在本研究中，我们采用一种鲁棒且低成本的触觉传感器BeadSight，作为精密预校准传感器的替代方案，用于操作任务的预训练方法。我们证明，即使使用BeadSight这样的低保真度传感器进行触觉预训练，也能提升模仿学习智能体在复杂操作任务中的表现。我们在USB线缆插接基准任务上验证了该方法，该任务先前是使用精度更高的GelSight传感器作为触觉输入进行预训练实现的。我们最佳的BeadSight预训练视觉-触觉智能体以70%的准确率完成了任务，而最佳GelSight预训练视觉-触觉智能体的准确率为85%，两者在推理阶段均仅使用视觉信息。