arXiv · 2025 · Withdrawn preprint

SpikeGrasp: A Benchmark for 6-DoF Grasp Pose Detection from Stereo Spike Streams

Zhuoheng Gao, Jiyao Zhang, Zhiyong Xie, Hao Dong, Zhaofei Yu, Rongmei Chen, Guozhang Chen, Tiejun Huang

SpikeGrasp infers 6-DoF robotic grasp poses directly from stereo spike streams via recurrent spiking networks without point-cloud reconstruction.

The scientific question

Can robots detect 6-DoF grasp poses directly from asynchronous spike streams?

Conventional robotic grasping relies on explicit 3D point-cloud reconstruction, a computationally heavy step absent from biological visuomotor pathways.

SpikeGrasp architecture for stereo spike-based grasp detection.
Figure 1 from the author-supplied manuscript; earlier arXiv version withdrawn ↗

The core idea

Point-Cloud-Free Visuomotor Spiking Pipeline

Mimicking biological visuomotor pathways, SpikeGrasp fuses binocular spike camera streams and uses a recurrent spiking network to iteratively refine 6-DoF grasp hypotheses end to end.

What we found

Data-Efficient Grasping in Clutter

Evaluated on a new large-scale synthetic stereo spike benchmark, SpikeGrasp outperforms point-cloud baselines in cluttered and textureless scenes with strong data efficiency.

The earlier arXiv preprint was withdrawn. The figure is from a different manuscript supplied by the author.

Continue the question

Related reading

Explore our research

Cite this work

Zhuoheng Gao, Jiyao Zhang, Zhiyong Xie, Hao Dong, Zhaofei Yu, Rongmei Chen, Guozhang Chen, Tiejun Huang. SpikeGrasp: A Benchmark for 6-DoF Grasp Pose Detection from Stereo Spike Streams. arXiv, 2025.