White Tiger‑VTouch is a newly released, open‑source visual‑tactile dataset that provides over 60,000 minutes of synchronized video, tactile sensor, and joint‑pose recordings from multiple robot platforms. Designed for cross‑body learning, it enables researchers to train models that transfer perception skills across humanoid, wheeled, and arm robots, dramatically expanding the data available for embodied AI.
Massive Scale for Cross‑Body Robot Perception
Unified Visual‑Tactile Streams
The collection aggregates high‑resolution video, precise tactile readings, and detailed joint‑pose information from diverse robots, including humanoid bodies, wheeled platforms, and articulated arms. All modalities are time‑aligned, creating a coherent multimodal experience that mirrors real‑world interaction.
Cross‑Body Learning Benefits
By aligning data across different hardware configurations, White Tiger‑VTouch enables models trained on one robot to apply perception skills to another without extensive retraining. This cross‑body capability reduces development time and broadens the applicability of learned behaviors.
Why Visual‑Tactile Integration Is Critical
Vision alone struggles with transparent objects, low‑light environments, and tasks requiring fine force control. Adding tactile feedback lets robots sense texture, compliance, and slip, bridging gaps that vision cannot fill. The dataset supports a shift toward integrated perception, essential for home assistance, precision manufacturing, and medical support.
Open‑Source Release and Community Impact
White Tiger‑VTouch is distributed under an open‑source license and hosted in a public repository. Unrestricted access encourages rapid iteration of perception algorithms, fosters collaboration, and creates a shared benchmark for evaluating multimodal models.
Research Opportunities with White Tiger‑VTouch
The dataset opens new avenues for embodied AI research, including:
- Domain‑Generalizable Manipulation: Train a single model to control diverse robot morphologies, leveraging cross‑body alignment to generalize grasp strategies.
- Sim‑to‑Real Transfer: Use extensive real‑world tactile recordings to narrow the reality gap for simulation‑trained policies.
- Fine‑Grained Material Recognition: Exploit tactile signatures to differentiate objects with similar visual appearance but distinct physical properties.
- Benchmarking Standards: Provide a common set of tasks and sensor configurations for fair comparison of perception algorithms.
Getting Started with the Dataset
Documentation includes ready‑to‑use scripts for loading synchronized video‑tactile streams into popular deep‑learning frameworks such as PyTorch. Early experiments demonstrate transformer‑based models that predict contact forces from visual cues alone, highlighting the dataset’s potential to improve perception even when tactile sensors are unavailable.
