VTouch++ Dataset Enhances Bimanual Manipulation with Vision-Based Tactile Sensing
Researchers have introduced VTOUCH++, a multimodal dataset designed to advance bimanual manipulation in robotics. The dataset leverages vision-based tactile sensing to capture high-fidelity physical interaction signals, addressing the lack of rich contact data in existing resources. It employs a matrix-style task design for systematic learning and automated data collection pipelines for scalability across real-world scenarios. Quantitative experiments on cross-modal retrieval and real-robot evaluations validate its effectiveness. The work demonstrates generalizable inference across multiple robots, policies, and tasks, marking a step forward in embodied intelligence for contact-rich manipulation.
Key facts
- VTOUCH++ dataset focuses on bimanual manipulation with vision-based tactile enhancement.
- Dataset provides high-fidelity physical interaction signals via vision-based tactile sensing.
- Matrix-style task design enables systematic learning.
- Automated data collection pipelines ensure scalability in real-world scenarios.
- Cross-modal retrieval and real-robot evaluations confirm dataset effectiveness.
- Generalizable inference demonstrated across multiple robots, policies, and tasks.
- Addresses limitations in existing datasets for contact-rich bimanual tasks.
- Published on arXiv under Computer Science > Robotics.
Entities
Institutions
- arXiv