Reproduce “Proprioceptive Image: An Image Representation of Proprioceptive Data from Quadruped Robots for Contact Estimation Learning” 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
Gabriel Fischer Abati, João Carlos Virgolino Soares, Giulio Turrisi, Victor Barasuol, Claudio Semini
Researchers at the Dynamic Legged Systems (DLS) lab, Istituto Italiano di Tecnologia, developed Proprioceptive Images (PIs), a 2D image representation for quadruped robot proprioceptive data that encodes both temporal dynamics and robot morphology. This representation, when used with a CNN-based model, improved contact estimation accuracy to 94.5% on real-world data, outperforming previous methods and demonstrating enhanced generalization to varied terrains with a significantly reduced input window.
Key insights
- Transforming proprioceptive time-series data into a structured 2D image, the Proprioceptive Image (PI), allows CNNs to effectively learn complex spatial and temporal features.
- Explicitly encoding robot morphology and temporal dynamics within the PI representation creates a richer feature space that captures inter-signal correlations and gait-dependent patterns.
- This image-based representation enables highly accurate predictions with significantly shorter temporal input windows, enhancing the efficiency and responsiveness of robotic learning systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.