Reproduce “Spark-0: Scaling Spatially Aligned Tactile Pretraining for Robot Dexterous Manipulation” 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
Tianxing Chen, Wenwei Lin, Chuqiao Lyu, Peicheng Xiang, Junyuan Tang, Tian Nian, Zijian Cai, Kehe Ye and 14 more
Spark-0 uses tactile demonstrations to train visual features to represent contact and conditions its action generator on touch when sensors are available. In a pretraining comparison, Ego Humanoid success without touch at deployment rises from 52.45% to 71.72% with 1,000 hours of tactile data. This supports retaining touch during pretraining, but does not isolate which route causes the gain.
Key insights
- Tactile data can serve as a training signal for visual representations, not only as an additional sensor input at inference; the contact-map supervision is removed for deployment.
- One model supports both vision-only and tactile-conditioned operation, replacing tactile tokens with learned placeholders when touch is unavailable.
- The clearest benefit from full-scale tactile pretraining appears in vision-only transfer; results are more modest on SparkArena-Sim, and the reported comparisons do not isolate individual tactile components.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.