Rico贾若童的博客

[ML] Graph Engineering

Graph Engineering for AI Agents: Increasing Certainty by Bounding Autonomy As models become more capable, the hard problem is no longer simply getting an agent to do something. It is getting the a...

Posted by Rico's Nerd Cluster on July 28, 2026

[ML] Attention Pooling

Attention Pooling Pooling converts a variable number of vectors into one fixed-size representation. For N point features, \[\mathbf{x}_1,\mathbf{x}_2,\ldots,\mathbf{x}_N,\] max pooling independe...

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

[ML] Continual Learning

LoRA, AdaBN

## 1. What continual learning actually means Suppose a detector is trained on Dataset 1. Later, Dataset 2 arrives. We want to train on Dataset 2 without keeping all of Dataset 1 in every training ...

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

[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

[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 June 8, 2026

Rico's Nerd Cluster