Reproduce “Contrastive Touch-to-Touch Pretraining” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Samanta Rodriguez, Yiming Dou, William van den Bogert, Miquel Oller, Kevin So, Andrew Owens, Nima Fazeli
Contrastive Touch-to-Touch Pretraining (CTTP) establishes a self-supervised framework to learn unified, sensor-agnostic representations from paired tactile data across distinct sensor modalities. This method enabled "zero-shot" transfer between sensors, improving cross-sensor classification accuracy significantly over baselines and leading to a 3.6x higher success rate in a real-world peg insertion task.
Key insights
- Contrastive learning can effectively align latent representations from distinct tactile sensor modalities, forming a unified, sensor-agnostic space.
- An optimal batch size during training (e.g., 128) is crucial for balancing the differentiation of objects with the alignment of sensor representations.
- The learned representations enable direct "zero-shot" transfer of models between different tactile sensors, overcoming the limitations of previous methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.