Rico's Nerd Cluster

「离开世界之前 一切都是过程」

[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...

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

BEV, Model Training Experience

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...

[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...

[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: ...

[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 ...

[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...

[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...

[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...

Preintegration Lio

1 - Time-Windowed State Estimate Imagine we have states at four times: \[x_1,\ x_2,\ x_3,\ x_4\] with odometry edges \[x_1-x_2,\quad x_2-x_3,\quad x_3-x_4\] and an additional edge \[x_1-x_3\]...

Fast Lio 2

TODO Remarks For observation update, FAST-LIO uses ikd-tree point-to-plane residuals instead of NDT residuals. The version above is a more teaching-style tightly coupled LIO view. FAST-LIO...