Skip to content

architecture.md

Latest commit

 

History

History
76 lines (55 loc) · 3.68 KB

architecture.md

File metadata and controls

76 lines (55 loc) · 3.68 KB

System Architecture | 系统架构

Overview | 概览

Q1 nano follows a Sim-to-Real pipeline: design the robot in CAD, simulate in MuJoCo, train walking policies with reinforcement learning, then deploy to real hardware.

┌─────────────┐    ┌──────────┐    ┌───────────┐    ┌──────────┐    ┌──────────┐
│ SolidWorks  │───▶│   URDF   │───▶│  MuJoCo   │───▶│ RL Train │───▶│  Deploy  │
│  (CAD)      │    │  Model   │    │ Simulation│    │ (Policy) │    │ (Real HW)│
└─────────────┘    └──────────┘    └───────────┘    └──────────┘    └──────────┘

Pipeline Stages | 流水线阶段

1. Hardware Design (CAD → URDF)

  • Input: Mechanical design in SolidWorks
  • Output: URDF model file for simulation
  • Location: hardware/solidworks/, hardware/step/, hardware/urdf/

The physical robot is designed in SolidWorks, exported as STEP for universal access, and converted to URDF for simulation compatibility.

2. Simulation (MuJoCo)

  • Input: URDF model
  • Output: Simulated environment for training
  • Location: simulation/mujoco/, simulation/configs/

MuJoCo provides a fast, accurate physics simulation. The URDF model is loaded into MuJoCo, where we can run thousands of episodes in parallel for RL training.

3. Reinforcement Learning Training

  • Input: MuJoCo simulation environment
  • Output: Trained walking policy (neural network weights)
  • Location: training/rl/, training/checkpoints/, training/scripts/

We use reinforcement learning (e.g., PPO) to train a walking policy. The agent learns to control joint positions to produce stable, natural-looking gait.

Key Training Details:

  • Observation space: joint positions, velocities, body orientation (IMU)
  • Action space: target joint positions for all servos
  • Reward: forward velocity + stability + energy efficiency + style penalties

4. Deployment (Sim-to-Real Transfer)

  • Input: Trained policy checkpoint
  • Output: Walking robot 🤖
  • Location: firmware/servo_control/, firmware/main_controller/

The trained policy is deployed to the real robot's microcontroller. The controller reads sensor data and outputs servo commands at ~50 Hz.

Sim-to-Real Challenges:

  • Domain randomization during training for robustness
  • Servo response delay compensation
  • Real-world surface friction variation

Hardware Architecture | 硬件架构

                    ┌─────────────────┐
                    │  Main Controller │
                    │  (ESP32/STM32)  │
                    └────────┬────────┘
                             │ Serial Bus
              ┌──────────────┼──────────────┐
              │              │              │
         ┌────┴────┐   ┌────┴────┐   ┌────┴────┐
         │ Servo 1 │   │ Servo 2 │   │ Servo N │
         │ (Hip L) │   │ (Hip R) │   │  (...)  │
         └─────────┘   └─────────┘   └─────────┘
  • 12 DOF (Degrees of Freedom): 6 per leg (hip yaw/roll/pitch, knee pitch, ankle pitch/roll)
  • Serial bus servos: All servos on a single bus, addressed by ID
  • Control loop: ~50 Hz policy inference → servo command