Building a cloud-edge-end three-layer decoupled multi-robot collaboration system from scratch — using large models for decision-making, edge computing for scheduling, and heterogeneous robots for pure execution.
How do you squeeze a GPU-dependent SpatialLM 3D large model into an embedded board with only 10 TOPS of compute? This article documents the practice of PC encoding + RDK X5 inference in a separated deployment architecture.
From natural language to autonomous action — using a three-tier large model system to enable a quadruped robot dog to understand human commands, remember environmental information, and autonomously navigate to execute tasks.
From five-stage pipeline to Cache subsystem to VLIW dual-issue architecture—building a complete ARMv3 processor from scratch on FPGA, understanding modern CPU microarchitecture design.
From overfitting to 91% accuracy—using ResNet152 + Inception v3 + Xception three-model feature fusion, solving 120 dog breed fine-grained classification through four iterative rounds.
Encoding piece ranks with resistor values lets the circuit board act as referee—from schematic to PCB layout, a complete electronic military chess design that requires no third player.
Using Silvaco ATLAS to simulate a complete PN junction diode from mesh definition to I-V characteristic extraction, understanding unidirectional conductivity from the physics level.
No CPU, no operating system—implementing Tetris on FPGA using pure Verilog hardware logic: rotation, collision detection, line clearing, and scoring for all seven tetrominoes with seven-segment display.
Building a complete FMCW radar signal processing simulation from scratch in MATLAB: from single-target ranging to multi-target 3D positioning, implementing the full algorithm chain for range, speed, and angle measurement.