RDK Challenge - 31/5/2k26 Venkat
#1
Posté 31 mai 2026 - 19:05
I'm L Venkata Subramanyam (Venkat), an engineering student from Gujarat, India, pursuing a minor in Robotics.
My technical background: Intermediate robotics developer with experience in ROS 2, SLAM, and computer vision
My robot is about (plan):
I plan to build the Real-Time CAD Extruder: an autonomous rover that drives around a room, building a live 3D structural mesh of the environment in a web browser as it moves. This is an embodied 3D modeling tool where the RDK X5 AI chip processes depth and geometry in real-time, and ROS 2 translates it into an extruded 3D mesh, basically the physical, driving version of a hand-gesture modeling tool.
Excited to join the challenge and learn with you all!
#RoboticsDreamKeeper #RDKX5
#2
Posté 31 mai 2026 - 19:05
#3
Posté 17 juin 2026 - 14:19
Just got it into my hands...
Let the build begin:flexed_biceps: ....
rn_image_picker_lib_temp_204a0d26-9b87-493a-afd6-aad79d09447a.jpg 2,69 Mo
0 téléchargement(s)
#4
Posté 17 juin 2026 - 15:26
#5
Posté 19 juin 2026 - 19:46
How can I connect my x5 with my laptop, I dont have monitor.
I tried connecting x5 with my friends monitor, it worked
#6
Posté 19 juin 2026 - 19:52
Bruh I dont have ethernet cable...
#7
Posté 19 juin 2026 - 21:39
root@ubuntu:~# ls
root@ubuntu:~# exit
logout
Connection to 10.223.173.9 closed.
C:\Users\vencu>ssh sunrise@10.223.173.9
sunrise@10.223.173.9's password:
Connection closed by 10.223.173.9 port 22
C:\Users\vencu>ssh sunrise@10.223.173.9
sunrise@10.223.173.9's password:
Permission denied, please try again.
C:\Users\vencu>ssh sunrise@10.223.173.9
sunrise@10.223.173.9's password:
Permission denied, please try again.
sunrise@10.223.173.9's password:
"
when i used r00t, it d0esnt have any access t0 files in sunrise user,
Als0 i didnt getany l0gin cmd when b00ted ( i used m0nit0r)
---
What d0 u thinkk is r00t issue here?
#8
Posté 19 juin 2026 - 21:43
"Last login: Sat Jun 20 04:22:33 2026 from 10.223.173.32
root@ubuntu:~# ls /userdata/
root@ubuntu:~# ls /opt/hobot/
model
root@ubuntu:~# ls /opt/hobot/model/x5/
basic
root@ubuntu:~# ls /app/
40pin_samples/ model/ pydev_demo/
cdev_demo/ multimedia_samples/
root@ubuntu:~#
"
& i have x5 connected to monitor now, i opened terminal, it has
"sunrise@ubuntu: "
as doalouge as of now.....
what to do?
ALso 'ls /userdata" in ssh didnt work...
#9
Posté 21 juin 2026 - 12:58
RDK X5 Board Bring-Up:
OS flashed successfully (Ubuntu)
SSH working from laptop
Phone camera streaming via IP Webcam
YOLO object detection running on BPU
Screenshots:
- Screenshot A: OS boot + SSH terminal
- Screenshot B: Phone camera stream
- Screenshot C: YOLO bounding boxes
GitHub: https://github.com/P...DK-X5-Challenge
#RoboticsDreamKeeper #RDKX5 #Stage1Complete
#10
Posté 22 juin 2026 - 04:33
#11
Posté 23 juin 2026 - 21:39
1. General BPU Hardware Limits: Ignoring the official model zoo for a second, what is the actual, physical ceiling of the Sunrise 5 BPU? In general, what is the heaviest model architecture you (or anyone) have actually managed to run on this chip before the RAM taps out or it crashes? I just want to know where the hardware's absolute breaking point is.
2. Architecture Check: Dual-Model RGB-D Point Cloud
To complete the multi-task requirement, here is the pipeline I'm planning to build:
Task 1 (BPU): YOLO-Seg to extract pixel-perfect 2D contours.
Task 2 (BPU): FCRN or FastDepth to generate a real-time depth map.
Task 3 (CPU): Take the YOLO mask, cross-reference it with the depth map, and stitch them together to output a local 3D .obj point cloud.
Does this dual-pipeline make sense for the RDK X5 to process smoothly, or is there a simpler/better combo to get Stage 3 checked off?
#12
Posté 23 juin 2026 - 22:09
the website:
https://d-robotics.g...40pin_define.md
is not working
with "Page Not Found
We could not find what you were looking for.
Please contact the owner of the site that linked you to the original URL and let them know their link is broken."
i went to that link from "https://d-robotics.g...3.5.0&p=RDK X5" s GPIO Interface section.
Thanks
#13
Posté 23 juin 2026 - 22:10
Thanks
#14
Posté 24 juin 2026 - 03:06
#15
Posté 24 juin 2026 - 03:59
YOLOv5-Seg 640×640 can reach around 126.64 FPS on RDK X5, so segmentation should not be the main bottleneck. The heavier part will likely be depth generation and CPU-side point cloud / OBJ processing.
For Stage 3, I'd suggest starting with:
YOLO-Seg -> object mask
Stereo depth -> depth map
CPU -> mask × depth -> local point cloud / OBJ
This should be simpler and more reliable than starting with FCRN/FastDepth. If you still want to try FCRN/FastDepth, you may need to export the model to ONNX and quantize it with the D-Robotics toolchain.
Quantization guide:
https://d-robotics.g...pment/overview/
#16
Posté 24 juin 2026 - 04:10
In the meantime, you can access the 40-pin GPIO documentation directly here:
https://d-robotics.g...le/40pin_define
The only difference is that the URL should not include the .md suffix.
Thanks for your feedback!
#17
Posté 24 juin 2026 - 11:46
#18
Posté 29 juin 2026 - 11:32
thanks f0r feedback.
I'm trying to use a mono. but n0t stere0 (because 0f time c0nstraints) & als0 is midas_v2.1 supprted bu rdk x5 b0ard? (Ill f0ll0w quantise...)
#19
Posté 29 juin 2026 - 12:14
#20
Posté 29 juin 2026 - 19:06
here is my Stage 2 submission:
https://github.com/P...ee/main/Stage 2
Please have a look whenever you are free...
Thanks!!!
1 utilisateur(s) li(sen)t ce sujet
0 members, 1 guests, 0 anonymous users









