Reproduce “Learn from Incomplete Tactile Data: Tactile Representation Learning with Masked Autoencoders” on OpenResearch

About this paper

Guanqun Cao, Jiaqi Jiang, Danushka Bollegala, Shan Luo

TacMAE, a method employing a masked autoencoder architecture, learns robust tactile representations by reconstructing missing information from incomplete sensor data. This approach demonstrated enhanced accuracy in tactile texture recognition, reaching 71.4% in zero-shot transfer and 85.8% with fine-tuning, and also improved object recognition on YCB objects with partial contact.

Key insights

• A high masking ratio (e.g., 70%) during training forces the model to infer substantial missing information, leading to the learning of more robust and generalizable tactile representations.
• Integrating a supervised classification head with the self-supervised masked reconstruction task significantly enhances the model's ability to learn discriminative features crucial for downstream recognition.
• The ability of the model to reconstruct tactile information from partial observations suggests that tactile signals on an object's surface contain sufficient redundancy for inference.

Summary by alphaXiv.