Reproduce “Near-sensor Computing for Rapid Visuotactile Perception” 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

Zhengying Zhu, Ruilin Zhang, Runze Hu, Chenxi Xiao

Researchers at ShanghaiTech University developed a near-sensor computing architecture for visuotactile perception, integrating a spectral Poisson solver directly into a sensor node to perform 3D depth map reconstruction. This system achieves a deterministic 0.211 ms latency for 128x128 reconstructions and reduces robotic protective reflex latency by 6.0-fold to 28.3 ms, with an estimated power consumption of 324 mW.

Key insights

  • Implementing a direct spectral Poisson solver in a fully pipelined hardware architecture enables deterministic, ultra-low latency 3D visuotactile reconstruction by avoiding iterative methods and host-based overheads.
  • Near-sensor integration of sensing, reconstruction, and decision-making dramatically reduces overall latency and variability, facilitating rapid robotic responses comparable to biological reflexes.
  • Utilizing 24-bit fixed-point arithmetic in the datapath provides an optimal balance between reconstruction accuracy and hardware resource efficiency for real-time operation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.