Embedded Files
Shutong Chen∗, Wenkai Zhang∗, Adnan Aijaz†, Miao Guo∗, Yansha Deng∗
∗Department of Engineering, King’s College London, London, UK
†Bristol Research and Innovation Laboratory, Toshiba Europe Ltd., Bristol, UK
Overview
We implement the first end-to-end Goal-oriented Communication testbed for Physical AI, which connects a physical robot and a commercial edge server through a 5G OpenAirInterface (OAI) network.
Main Contribution 1:
We implement onboard semantic extraction at a Piper robot equipped with an RGB-D camera and a 5G modem, language model inference, digital twin reconstruction, and trajectory replanning at the edge server deployed on an NVIDIA Jetson AGX Orin, and 5G OAI network between them.
Main Contribution 2:
We experimentally characterise the data sizes of different transmitted semantic representations and full-stack transmission latency, quantifying how different data types affect sensing, communication, and computation performance.
Main Contribution 3:
We conduct extensive experiments over representative Physical AI tasks and report the latency breakdown across the complete S3C loop, together with the task success rate, showing that our GoC design substantially reduce task completion time while improving the task success rate.
Experiment Overview: Physical AI Fault Detection and Recovery
Case 1 : Grasp Failure
Baseline Design
Baseline Design
IMG_4515.movOur GoC Design
Our GoC Design
IMG_4517.movCase 2 : Object Drop
Baseline Design
Baseline Design
IMG_4520.movOur GoC Design
Our GoC Design
IMG_4523.movCase 3 : Incorrect Placement
Baseline Design
Baseline Design
IMG_4525.movOur GoC Design
Our GoC Design
IMG_4527.movExperiment Overview: Fruit Packing
TODO
Experiment Overview: Cube Stacking
TODO
Page updated
Google Sites
Report abuse