Rico贾若童的博客

[BEV] BEV Introduction

Understanding L2 and L2+ Perception: Why the Real Challenge Is System Design, Not Just Better Detection When people talk about assisted driving, they often jump straight to compute power, percepti...

Posted by Rico's Nerd Cluster on April 2, 2026

MPPI And Motion Control On My Diff Drive

Remote Claude Code Setup

Motor Control So, tested on carpet, I think it’s about not having (+/-) on the spot that makes it motor controller - aware that the robot turns better using something like: left = ±u righ...

Posted by Rico's Nerd Cluster on March 25, 2026

AI Agent Setup

Remote Claude Code Setup, Claude Session

Remote Claude Code Setup Bring up tailscale: sudo tailscale up Bring up code-server: code-server --bind-addr $(tailscale ip -4):8081 To see password for access cat ~/.config/code-server/con...

Posted by Rico's Nerd Cluster on March 22, 2026

Computer Vision - Charuco Board and Phone Pose Detection

Mission The goal is to generate an object ground-truth pose using an RGB-D camera, a ChArUco board, and a known object CAD model or object dimensions. The ChArUco board gives us a stable metric c...

Posted by Rico's Nerd Cluster on March 18, 2026

Computer Vision - Charuco Board Detection

ChArUco Pose Estimation A ChArUco board is basically a chessboard combined with ArUco markers. An ArUco marker is a square fiducial marker with a binary ID pattern inside it: 1 2 3 4 5 6 marker ...

Posted by Rico's Nerd Cluster on March 16, 2026

Computer Vision - Aruco Detection

Aruco Detection Aruco detection has three jobs: 1 2 3 1. Detection: find square marker candidates in the image. 2. Decoding: read the binary ID inside the square. 3. Pose estimation: use the four...

Posted by Rico's Nerd Cluster on March 15, 2026

Computer Vision - Pose Estimate Pipeline

Building a 2D-to-6D Object Pose Pipeline with RF-DETR and FoundationPose I built a small pose-estimation pipeline that combines RF-DETR and FoundationPose. RF-DETR solves the 2D detection problem:...

Posted by Rico's Nerd Cluster on March 13, 2026

[CUDA - 6] CUDA Inference Speedups

inference mode, jit tracing, cuda graphs, tensor rt, optimize_for_inference

Introduction I built a RF-DETR + FoundationPose pipeline. the clean optimization stack is: Make inference measurable. Log per-stage latency: detector, register, track, publish, visualize. Us...

Posted by Rico's Nerd Cluster on March 12, 2026

DETR

Embedding, Encoding An embedding is a learnable vector, part of an embedding table A, which is similar to a fully connected layer y = x @ W.T + b. However, an embedding usually does: 1 one_hot(in...

Posted by Rico's Nerd Cluster on March 11, 2026

ViT

ViT in short: Image -> 16x16 patches [16 x 16 x 3] -> project to [256/768] token embedding TODO

Posted by Rico's Nerd Cluster on March 8, 2026

Rico's Nerd Cluster