TouchNet

TouchNet is the tactile dataset project described by Eric Li Zhiqiang / 李志强 in 从会跳舞到有感知，触觉是机器人通往智能的门票吗？｜ S10E19. Yimu Technology / 一目科技 frames it as an ImageNet-like effort for Tactile Sensing: a shared data resource that could make touch models easier to train, compare, and integrate into robot systems.

The source treats TouchNet as necessary because tactile AI lacks the historical data inheritance that computer vision received from large image datasets. The challenge is not only volume. Touch data also needs physical definitions for softness, hardness, roughness, smoothness, friction, and slip, plus alignment across tactile, visual, audio, and temporal streams.

Key Claims

• Tactile AI is data-poor compared with vision and language, so shared datasets may be a field-level accelerator.
• Touch labels are harder than image labels because many tactile concepts require physical thresholds and continuous time-series evidence.
• TouchNet would support Tactile Transformer Encoder training and multimodal alignment with visual robot inputs.
• The dataset sits inside a broader Embodied Data Pyramid rather than replacing real deployment, simulation, or video pretraining.

Connections

• ImageNet — analogy for a dataset that accelerates a field.
• Yimu Technology / 一目科技 and Eric Li Zhiqiang / 李志强 — organization and speaker attached to the project.
• Tactile Sensing, Optical Tactile Sensing, and Tactile Transformer Encoder — data domain, sensor route, and model interface.
• Embodied Data Pyramid, Real Robot Data Strategy, and Robotics Simulation Evaluation — broader robot-data strategy concepts.