iCub¶

Repository for the resources related to the Nina robot :

• Source files that are used on all the iCub related computers
• Source files that are used on some specific iCub related computers
• Documents that help to run a demo, conduct an experiment
• Status since 2025's Nina_update

Documentation on mical-008:¶

• has CUDA 12.8
• python3.12 is available system-wide
• yarp is available system-wide
• import my_yarp to access local or remote yarp server, from python
• torch '2.6.0+cu126' from /opt/python-env/
• can run several yarp-ed services: YOLO, vllm... see Mical-008: services for details

• needing conda ? export PATH="/opt/miniforge3/bin:$PATH"

• shared directories :
  • /opt/miniforge3/
  • /localdata/robotology/
  • /opt/python-env/

Please note:

• mical-008 is not in the same VLAN as desktops or gpu3. It's on the Mical VLAN, like Nina and Furhat.
• X: will not be propagated by ssh-interne. Ok inside Mical VLAN though

Mical-008 services¶

To help usages on both robots (or without), remotely, with yarp, without taking care of the venv or conda environments... Some services are created/runnable from your home directory:

• service_yarp_vllm: generic output or scenario-specific output (if prefixed with TASK=Unanimo, TASK=Rephrase...)
• service_yarp_yolo: will grab its images from a remote webcam server
• service_yarp_furhat_tts: will generate speech/articulation for AD or IZ speaker, and send them to furhat
• apache docker: can serve files from /localdata/httpd/html on local port
• service_whisper

Mical-008: to do list¶

• DONE: install and testTTS <-- forced to have a python3.10
• DONE: test GPU usage of whisper
• DONE: implement "services", easy to run/terminate, even remotely and with yarpmanager
• DONE: v1 have vLLM run as a YARP service (yarp rpc)
• DONE: find a solution to send local webcam video stream/image to GPU
• DONE: have yolo inference on GPU server, instead of local Windows PC. Can use YOLO "COCOA" model or GIPSA "card" model. Can use table-top camera, or Furhat front camera.
• train yolo on the new GPU, with bigger card sets and more hand data
• load test, with everything at the same time
• test omni2