---
license: cc-by-4.0
tags:
  - robotics
  - mlp
  - numpy
  - obstacle-avoidance
  - simulation
pretty_name: Sim Driving MLP (NumPy)
---

# Sim Driving MLP — 279-parameter obstacle-avoidance policy (NumPy)

A deliberately tiny neural network: **4 ultrasonic distances in, one driving command
out.** 279 parameters, float32, trained and verified entirely inside our robot
simulation studio. Small enough to read, small enough to run on an MCU-class device.

**Trained on simulation data from a virtual map — no real-world or customer data.**

## Architecture

![Architecture: 4 → 16 → 8 → 6](assets/01_architecture.png)

| part | meaning |
|------|---------|
| input (4) | ultrasonic distances: **F**ront, **L**eft, **R**ight, **B**ack (normalized with the included `norm_mean` / `norm_std`) |
| output (5 + 1) | 5 command logits — `FWD / LEFT / RIGHT / STOP / BACK` — plus 1 turning-angle regression head |

## Metrics (validation, virtual map)

| metric | value |
|--------|-------|
| command accuracy | **97.2 %** (best epoch 529 / 572) |
| turning-angle MAE | ~6.9° |

The `.npz` also embeds `meta_json`: the full 572-epoch training history
(per-class accuracy, loss, angle MAE per epoch), so the training curve is inspectable.
The plot below is drawn directly from that embedded history:

![Training history](assets/02_training_curve.png)

## Load and run (NumPy only, no framework)

```python
import numpy as np

d = np.load("sim_driving_mlp.npz")
x = np.array([[120.0, 45.0, 200.0, 300.0]])  # F, L, R, B distances (mm)
x = (x - d["norm_mean"]) / d["norm_std"]

h = np.maximum(x @ d["W0"] + d["b0"], 0)
h = np.maximum(h @ d["W1"] + d["b1"], 0)
y = h @ d["W2"] + d["b2"]

cmd = ["FWD", "LEFT", "RIGHT", "STOP", "BACK"][int(np.argmax(y[0, :5]))]
angle = float(y[0, 5])
print(cmd, angle)
```

## The simulator and the target robot

The studio stage where the model's four inputs are defined: ultrasonic sensors
F/L/R/B with datasheet-based noise models, checked by the built-in self QA/QC
checklist. On the right, the local LLM explains a warning from that checklist,
citing the stage report as its basis.

![Sensor definition stage in the simulation studio](assets/05_sim_studio_sensors.png)

A driving run on the 8 m × 8 m virtual map used for data collection
(green: ultrasonic rays from the robot):

![Driving run on the virtual map](assets/06_virtual_map_run.png)

And this exact model running in the studio's 3D evaluation stage. The left panel
shows the live inference at the current step: the four sensor inputs (F/L/R/B, mm),
the softmax over the five commands with FORWARD selected, and the angle head:

![The model driving in the 3D evaluation stage, with live inference panel](assets/07_nn_eval_3d.png)

The target hardware: a tracked test robot with the ultrasonic sensors mounted on
the hand, the same F-channel placement the simulator reproduces.

![Tracked test robot with hand-mounted ultrasonic sensors](assets/04_real_robot.jpg)

## Where this fits: our 4-layer stack

![NCDTech 4-layer stack](assets/03_four_layer_stack.png)

*(Diagram is in Korean; it is the same figure used on our website, demo video, and
companion dataset.)*

This model is a **layer-2 artifact** of our stack — an edge neural network verified
through the 8-stage physics simulation workflow of our robot simulation studio.
Every stage of that workflow runs self QA/QC (layer 3) and reports through an
on-premise conversational LLM (layer 4); the record formats those layers produce
are shown in our companion dataset:
[NCDTech/human-gated-qaqc-knowledge-example](https://huggingface.co/datasets/NCDTech/human-gated-qaqc-knowledge-example)

## 한국어

**초음파 4방향 거리(전·좌·우·후)를 넣으면 주행 명령이 나오는 279 파라미터짜리 작은
신경망**입니다. 저희 로봇 시뮬레이션 스튜디오의 8단계 워크플로(데이터 수집 → 학습 →
평가 → 물리 시뮬레이션 검증)를 통과한 계층 2(엣지 신경망) 산출물이며, **가상 맵
시뮬레이션 데이터로만 학습**했습니다 — 실데이터·고객 데이터 없음.

출력은 주행 명령 5클래스(전진/좌회전/우회전/정지/후진) + 회전각 회귀 1개.
검증 정확도 97.2 %, 각도 오차 약 6.9°. 파일 안에 정규화 통계와 572 에포크 학습
이력 전체가 함께 들어 있어 학습 곡선을 그대로 확인할 수 있습니다.

Learn more: https://huggingface.co/NCDTech · https://www.ncdtech.org