Rico贾若童的博客

[Robotics] Two Paths from PyTorch to Fast GPU Inference

TensorRT, Torch Compile

torch.compile and TensorRT A PyTorch model can reach an NVIDIA GPU through several optimization paths. Two important choices are: compile PyTorch execution with torch.compile and TorchInductor...

Posted by Rico's Nerd Cluster on July 19, 2026

[ML] From GPU Kernel to Machine Code

Triton, CUDA, PTX, and SASS

Overview When you write a GPU operation, several different languages and compilation stages may sit between your source code and the instructions executed by the GPU. The short version is: 1 2 3...

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

[Robotics] 3D-shape-detection-from-point-clouds

CenterPoint

BEV BEV is full 7 dof representation: (x,y,z,l, w, h, yaw). Its scene is a heatmap on a top down grid, and where it localizes its peak. Here, (x,y,z) is a BEV pillar, which is the 2.5D representa...

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

[Robotics] Sonar Noise Analysis

phasor, rayleigh distribution, exponential distribution, gamma distribution, TVG simulation

Sonar noise have two independent characteristics: Intensity noise: usually multiplicative Gamma speckle. Range noise: usually additive, spatially correlated residuals. Time varying gain : t...

Posted by Rico's Nerd Cluster on July 10, 2026

[Robotics] Time Series Analysis Metrics

MAD

1 - MAD There are two different definitions of MAD. In machine learning papers, you have to check which one they mean. 1- 1 Mean Absolute Deviation Given measurements \[x = [3,4,5,8]\] Mean: ...

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

[Product Ideas] Agentic Loop and Moving Fast

Agentic Loops Study: Move Fast and Be Responsible The first step in building agentic products is knowing what to build. Survey data is useful, but it should not be treated as a direct instruction ...

Posted by Rico's Nerd Cluster on June 20, 2026

[Career Advice] REPOST - What 16 years of software engineering taught me about growth

This article was originally posted here In service companies, ignoring documentation meant chaotic client handoffs. In product companies, skipping community meant slower growth. In my startup, avo...

Posted by Rico's Nerd Cluster on May 29, 2026

[Robotics] Retrain RF DETR

RF-DETR Fine-Tuning for a Custom Class When fine-tuning RF-DETR on a custom object, the usual workflow is: 1 2 3 pretrained RF-DETR backbone/features -> replace or reinitialize the classificat...

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

[Robotics] From Market-Ready ROVs to Low-Cost AUVs: Building Practical Underwater Autonomy

This is a summary of From Market-Ready ROVs to Low-Cost AUVs BlueROV2 is a small 6-DoF ROV ROV platform that can operate at depths up to 100 meters, weighs under 12 kg, and includes a tether of up...

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

NDT-Tightly-Coupled-Lio

iEKF

Loosely vs Tightly Coupled LIO LiDAR-inertial odometry combines LiDAR geometry with IMU motion prediction. At a high level, there are two common ways to do this: loosely coupled fusion and tightly...

Posted by Rico's Nerd Cluster on May 3, 2026

Rico's Nerd Cluster