Voice-Controlled Robot Arm

A full-stack embodied AI system — voice in, physical action out — running entirely offline on consumer hardware.

中文

---

Overview

Layer	Implementation
Hear	Faster-Whisper, local Chinese speech recognition
Think	DeepSeek-R1-1.5B + QLoRA fine-tune, natural language → JSON
See	YOLOv8s object detection + homography hand-eye calibration
Move	D-H inverse kinematics + S-Curve trajectory, ESP32 PWM

Total hardware cost ¥317 (~$45 USD). Requires an NVIDIA GPU for LLM inference (RTX 3060 6GB recommended, <4GB VRAM at runtime, <200ms latency).

---

Architecture

flowchart TD
    MIC["🎤 Microphone"] --> STT["Faster-Whisper<br/>Chinese speech recognition"]
    STT --> RULE{"Regex engine<br/>Simple command match"}
    RULE -- "Hit" --> ACT["JSON action"]
    RULE -- "Miss (has object name)" --> LLM["DeepSeek-R1-1.5B<br/>QLoRA FP16<br/>Natural language → JSON"]
    LLM --> ACT
    ACT --> VIS["YOLOv8s + Homography<br/>Object detection · hand-eye calibration<br/>Pixel coords → robot coords mm"]
    VIS --> MOT["arm_main.py<br/>D-H IK + S-Curve"]
    MOT --> ESP["ESP32 PWM → Servos"]

---

Bill of Materials

Total: ¥317 (~$45 USD)

#	Item	Spec	Qty	Unit	Total
1	3D-printed robot arm kit	Acrylic/PLA structural parts	1	¥71	¥71
2	ESP32 dev board	Dual-core MCU, WiFi + BT	1	¥19	¥19
3	ESP32 accessories	Connectors / expansion board	1	¥5	¥5
4	USB industrial camera	Plug-and-play, wide-angle, 1280×720	1	¥61	¥61
5	Digital servo MG996R	Metal gear, high torque	5	¥27	¥133
6	Regulated power supply	6V 6A, servo-dedicated	1	¥29	¥29

Wiring

• ESP32 pins: X→14, Y→4, Z→5, B→18, Gripper→23
• Power: servos and ESP32 on separate supplies (external 6V/6A) to prevent inrush surge
• Camera: USB, mounted in front of the arm covering the full work surface
• Serial: USB to ESP32, default port COM3, override with ROBOT_PORT env var

---

Installation

1. Flash Firmware

Arduino IDE 2.x, board: "ESP32 Dev Module". Open main.ino, select the correct port, click Upload.

2. Python Environment

Python 3.10+, CUDA 11.8 or 12.x.

# Install the correct CUDA build of PyTorch from pytorch.org first, then:
pip install -r requirements.txt

3. Configure

All tunables are in config.py and support environment variable overrides — no code changes needed:

ROBOT_PORT=COM5               python voice_main.py  # change serial port
LLM_MODEL_PATH=D:\models\lora python voice_main.py  # change LLM path
YOLO_MODEL_PATH=runs/best.pt  python voice_main.py  # change YOLO path

4. Models

Speech (Whisper): the base model is downloaded automatically on first run.

Vision (YOLO): train your own detector — 50 labelled images is enough for transfer learning:

yolo detect train model=yolov8s.pt data=data.yaml epochs=100 imgsz=640
# Output: runs/detect/train/weights/best.pt → copy to project root

LLM: fine-tune DeepSeek-R1-1.5B or Qwen1.5-1.8B with QLoRA. See TRAINING.md for the complete guide.

Training data format (Alpaca):

{
  "instruction": "lift the pencil sharpener 5cm",
  "input": "",
  "system": "You are a robot arm JSON converter...",
  "output": "[{\"action\": \"lift\", \"target\": \"part\", \"height\": 50}]"
}

---

Quick Start

python voice_main.py

On startup the system loads in order: serial port → YOLO → Whisper → LLM → camera window.

Keyboard Shortcuts

Key	Function
SPACE (hold)	Record audio; release to transcribe and execute
C	Toggle hand-eye calibration mode
R	Manual reset to home position
O	Force open gripper
Q	Quit

---

Voice Commands

Speak natural Chinese. No special syntax required.

Pick and transport (requires visual detection)

"把削笔刀抓起来"   — pick up the pencil sharpener
"抓住那个盒子"     — grab that box
"把削笔刀抬起5厘米" — lift the pencil sharpener 5cm
"将零件举高10公分"  — raise the part 10cm

Precise directional movement

"向上三厘米"      → Z +30mm
"向左移动四毫米"   → Y +4mm
"往前伸10厘米"    → X +100mm

Fuzzy movement (no explicit distance, defaults to 5cm per config.DEFAULT_MOVE_MM)

"向左"  "抬起"  "往下"

Gestures and state commands

"点头"  — nod: oscillate Z ×3 (±3cm)
"摇头"  — shake head: oscillate Y ×3 (±3cm)
"放下"  — lower to table height (Z=-15mm) and release
"复位"  — return to home position [120, 0, 60] mm
"松开"  — open gripper without moving

Speech compatibility: built-in homophone correction for common Whisper mishearings, e.g. "零米"→"厘米", "小笔刀"→"削笔刀", "电头"→"点头".

---

Hand-Eye Calibration

Recalibrate whenever the camera is moved. Press C to enter calibration mode, then click 4 corner points in order:

P1 (top-left)     ↔  robot coords (90,  90)
P2 (top-right)    ↔  robot coords (200,  90)
P3 (bottom-right) ↔  robot coords (200, -90)
P4 (bottom-left)  ↔  robot coords (90, -90)

The homography matrix updates instantly after the 4th click. No restart needed.

---

Troubleshooting

Symptom	Cause	Fix
SPACE does nothing	Camera window not focused	Click the camera window first
Garbled recognition	Mic noise / speaking too fast	Quiet environment, moderate pace; hold SPACE 0.5s before speaking
"Target not found"	YOLO didn't detect the object	Adjust lighting/angle; verify object is in training classes
Pick position offset	Camera was moved	Press C and redo 4-point calibration
Serial connection failed	ESP32 not plugged in / wrong port	Check device manager; set ROBOT_PORT env var
Violent shaking on startup	5-servo simultaneous inrush	Firmware staggers power-on; if it persists, check PSU capacity

---

Technical Notes

Key engineering problems solved during development.

D-H Inverse Kinematics The 130mm L4 link causes ~40° path deviation with geometric IK during horizontal moves. Solved by Scipy SLSQP numerical optimization with a Pitch=-90° constraint (end-effector always perpendicular to the table), eliminating the nonlinear offset entirely.

S-Curve + Multi-Layer Damping MG996R servos vibrate badly under a long lever arm. Five-layer damping pipeline: tilt correction → moving-average filter (deque) → speed cap → EMA damping → dead-zone filter.

Dual-Channel Parse Architecture Simple commands (release/reset/directional moves) bypass the LLM entirely via a regex engine (microseconds). Only complex commands containing object names reach the LLM (<200ms). This prevents the common failure mode where "move down 3cm" gets misclassified as a lift action.

Pre-filling to Skip Chain-of-Thought DeepSeek-R1 outputs a <think>...</think> chain-of-thought by default. Appending <｜Assistant｜> as a pre-fill token forces the model to skip the thinking phase and emit JSON directly, achieving 100% format compliance.

Whisper Anti-Hallucination Three defences, all encapsulated in RobotEar.get_text(): silence trimming + duration guards; condition_on_previous_text=False; repeated-phrase regex dedup (removes "向右向右向右..." loops). All thresholds are tunable via config.py.

Engineering Pitfall: System Prompt Alignment The system prompt at inference must exactly match the one used during fine-tuning. Any mismatch causes output drift (e.g., outputting 500mm instead of 50mm). A warning comment is included in the source.

---

LLM Training

~500 domain-specific samples, QLoRA fine-tune of DeepSeek-R1-1.5B, loss converged to 0.0519, format error rate 0%.

See TRAINING.md for the full guide: QLoRA hyperparameter config, GGUF vs Transformers comparison, pre-filling inference details, and experiment results.

---

Project Structure

robot_arm/
├── README.md          Chinese documentation
├── README_EN.md       This file
├── TRAINING.md        LLM LoRA fine-tuning research notes
├── requirements.txt   Python dependencies
├── config.py          All tunables: hardware, motion, audio & gesture constants
│
├── main.ino           ESP32 firmware, LEDC PWM servo control
├── arm_main.py        Kinematics core: D-H IK + S-Curve trajectory
├── whisper_main.py    Full ASR pipeline: silence trim → transcribe → post-process
└── voice_main.py      Main app: voice → LLM → vision → motion

---

Key Specs

Metric	Value
Hardware cost	¥317 (~$45 USD)
GPU requirement	RTX 3060 6GB (<4GB VRAM at runtime)
Inference latency	<200ms (LLM), <50ms (rule engine)
Training samples	~500
Format error rate	0%
Operation mode	Fully offline