Learning with Less: Optimizing Tactile Sensor Configurations for Dexterous Manipulation

TL;DR

This work tackles the practical challenge of tactile sensor deployment for dexterous manipulation by systematically evaluating how sensor quantity and placement affect DRL-based control on the Shadow Hand. It introduces a 21-sensor configuration that preserves about $93\%$ of full-performance while reducing sensor count by approximately $77\%$, and pairs this with a multi-factor regression model that predicts task success across arbitrary layouts with high accuracy ($\sim96\%$) and good generalization. The approach includes a hybrid prediction framework that combines correlation-based sensor importance with linear regression, and demonstrates robustness to interference and transferability to unseen tasks. Collectively, the results offer a scalable, cost-effective pathway for deploying tactile sensing in real-world dexterous manipulation systems while maintaining performance and robustness across tasks.

Abstract

Tactile sensing is critical for learning-based robotic dexterous manipulation, enabling real-time force perception, slip detection, and grip adjustments during interactions. While full-hand sensor arrays provide precise control, their deployment is limited by high costs, complex integration, and significant computational demands. Practical constraints, including limited space and the complexity of the wiring, further restrict the use of the entire sensor. Consequently, optimizing sensor configurations to achieve efficient coverage and good performance using fewer sensors remains a significant and open research challenge.In this work, we investigate the influence of tactile sensor quantity and placement on a robotic hand for dexterous manipulation tasks. Through systematic analysis of various sensor configurations, an optimized layout with only 21 sensors is identified, achieving over 93% of the task success rate relative to full-hand coverage (92 sensors). This configuration reduces the sensor count by 77% and offers a considerable reduction in integration costs, demonstrating a cost-effective yet high-performing tactile sensing strategy. Additionally, we develop a multi-factor regression model to predict task success rate under arbitrary sensor configurations. The model achieves strong generalization, with an average prediction error of 3.12% on unseen manipulation tasks. These results offer a scalable framework for deploying tactile sensing in real-world robotic manipulation systems.

Figures (5)

• Figure : Fig. 1. Overview of the sensor configuration optimization pipeline. First, a sensor ablation study is conducted using a DRL agent in robot hand manipulation tasks to evaluate performance across configurations. Second, a regression model is trained to predict task performance with 96% accuracy and strong generalization across tasks. Third, the optimized configuration and model are validated on five manipulation tasks to assess generalization and robustness under task transfer. Finally, performance–cost analysis reveals that using only 21 sensors (22% of the full set) achieves at least 93% of the original performance, offering the best performance-to-cost trade-off.
• Figure :
• Figure : Fig. 3(a). The result of the quantity study. The configuration $\mathbf{A_{21}}$ with the least sensors achieves 93% of task success rate compared to $\mathbf{A_{92}}$. (b) shows the result of the placement study. Configurations with closed success rates are shown in the same color.
• Figure : Fig. 4(a). The detailed coefficients of the fine-tuned prediction model reflect the importance and influence of each sensor on the final performance. Fig. 4(b). The placements of the optimized 21-sensor configuration.
• Figure : Fig. 5. The result of the interference study, and the red dot line represents the success rate baseline, which is without a sensor.