Tactus: Open-Vocabulary Object Recognition from Low-Cost Pressure Arrays
Summary:A new arXiv paper introduces Tactus, an open model for open-vocabulary object recognition from low-cost resistive pressure arrays. It achieves 0.771 top-1 accuracy on the STAG benchmark (27 objects), surpassing the supervised closed-set CNN (0.76) without a trained classifier head. The recipe is small-data: 187 training recordings, masked-autoencoder pretraining on 144k unlabeled frames, and the sensor's calibration affine. Failures are reported with equal precision, and weights, code, and memor
Overview
Tactile sensing is a critical modality in robotics, but tactile representation learning has focused on optical sensors that image a deforming gel. Tactus addresses this gap by performing open-vocabulary object recognition directly from pressure data, answering text queries without a trained classifier head.
Technical Highlights
- Open-Vocabulary Recognition: Tactus matches arbitrary text queries, achieving 0.771 top-1 accuracy (top-3 0.935) on the STAG benchmark, surpassing the supervised closed-set CNN (0.76).
- Small-Data Recipe: Only 187 training recordings, combined with masked-autoencoder pretraining on 144k unlabeled frames and the sensor's calibration affine, yield high performance. The calibration affine recovered more accuracy than all architecture changes combined.
- Robustness Analysis: Errors concentrate in a few contact-ambiguous classes, are uncorrelated with text-target geometry (Spearman rho ≤ 0.05), and survive paraphrased and bare-name queries. Two diverse frames recover 89% of eight-frame accuracy.
- Failure Reporting: Cross-sensor pretraining pooling gave no gain, vision co-training degraded touch, and a mis-normalized input pipeline silently discarded 97% of the sensor's dynamic range. These findings provide important warnings for future research.
Industry Impact and Developer Advice
Tactus demonstrates the potential of low-cost tactile sensors in open-vocabulary recognition, offering new insights for robotic tactile perception. Developers can benefit from its small-data training strategy and the importance of calibration affine, while avoiding pitfalls like cross-sensor and vision co-training. Weights and code are open-sourced for reproducibility.
Source
Based on arXiv paper 2608.04043 (released August 6, 2026).
— END —Source: ArXiv Machine Learning (cs.LG) (2026-08-06)
Tags: #Touch #Open Glossary #Pressure array #Self-supervised learning #Robot
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments