Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

A Segmented Robot Grasping Perception Neural Network for Edge AI

Created by
  • Haebom

Author

Casper Brocheler, Thomas Vroom, Derrick Timmermans, Alan van den Akker, Guangzhi Tang, Charalampos S. Kouzinopoulos, Rico Mockel

Outline

This paper presents a robotic gripping technology capable of stably grasping and manipulating objects of various shapes, sizes, and orientations. Specifically, we implement Heatmap-Guided Grasp Detection, an end-to-end framework for detecting six-degrees-of-freedom (6-DoF) grip poses using a deep neural network that learns rich, abstract representations of objects, on a GAP9 RISC-V system-on-chip (SoC). We optimize the model using hardware-aware techniques (input dimensionality reduction, model partitioning, and quantization) and verify the feasibility of full on-chip inference using the GraspNet-1Billion benchmark, highlighting the potential for real-time autonomous manipulation using low-power MCUs.

Takeaways, Limitations

•
Takeaways:
◦
Demonstrating real-time robot gripping on a low-power MCU.
◦
Presenting the possibility of efficient model optimization through hardware-aware techniques.
◦
Expanding the potential for real-time autonomous operation through low-latency, low-power inference on edge devices.
•
Limitations:
◦
Experimental results limited to a specific hardware platform called GAP9 SoC.
◦
Only performance evaluations for the GraspNet-1Billion benchmark are presented; performance on other benchmarks or in real-world environments has not been verified.
◦
Further research is needed to determine the generalizability of the hardware recognition techniques used.
👍
Made with Slashpage