EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video

Researchers introduce EgoTactile, a new benchmark and AI framework for estimating hand grasp pressure from egocentric video without intrusive hardware sensors. The work combines vision-based deep learning with diffusion models to infer tactile information for VR and robotic applications, achieving strong generalization to real-world scenarios.

EgoTactile addresses a fundamental challenge in embodied AI: inferring tactile information from visual data alone. Current robotic and VR systems struggle with grasp estimation because dense pressure sensing typically requires expensive or intrusive hardware that limits practical deployment. This research shifts the paradigm by extracting tactile signals purely from egocentric video, reducing hardware barriers while improving system accessibility.

The technical contribution extends beyond simple baseline models. While EgoPressureFormer serves as a discriminative baseline, the EgoPressureDiff diffusion-based framework represents a meaningful advance by explicitly modeling uncertainty in partial observations. By leveraging pre-trained video diffusion models and introducing a Physically-Informed Feature Rectification layer, the approach bridges visual and physical domains—resolving the ambiguity inherent when inferring contact patterns from incomplete visual information.

For roboticists and VR developers, this work reduces implementation complexity and cost. Robotic manipulators can achieve better grasp control without specialized sensor arrays, while VR systems can provide richer haptic feedback through computationally inferred pressure estimates. The bare-hand transfer subset demonstrates practical applicability beyond controlled laboratory settings, suggesting deployment potential in consumer-grade devices.

The research validates that pre-trained foundation models contain sufficient world knowledge to infer physical properties not explicitly represented in training data. This pattern—extracting hidden information from multimodal models—increasingly defines modern AI development. Future work likely focuses on expanding to more object categories and real-time performance optimization for edge deployment in robotic systems and consumer hardware.

• →EgoTactile enables grasp pressure estimation from video without hardware sensors, reducing deployment barriers for robotics and VR applications
• →Diffusion-based architecture outperforms discriminative baselines by explicitly modeling uncertainty in partial visual observations
• →Physically-informed constraints bridge visual and tactile domains, resolving ambiguities in contact pattern inference
• →Strong generalization to uncontrolled scenarios demonstrates practical applicability beyond laboratory benchmarks
• →Pre-trained video models contain sufficient world knowledge to infer non-visual physical properties through semantic feature rectification