3WE Robot Platform

AI-First Open Infrastructure for Embodied Robotics

An open-source robot platform for Embodied AI research — a consistent Python API across simulation and real hardware, with <$500 reproducible hardware.

Status: Active development. Mock backend is functional. Hardware builds and Gazebo integration are being validated — no completed physical prototype video yet.

A note on Sim2Real: The Python API stays the same when you switch backends, but sim-to-real transfer in practice still requires task-specific tuning (sensor models, noise, dynamics). We provide presets for common scenarios; we don't claim the gap away. See Sim-to-Real Gap for what works, what doesn't, and where you'll likely need to do downstream work.

Autonomous navigation in office_v2 scene — 360° LiDAR, real-time obstacle avoidance, 4 lines of Python

Note: The animation above is a concept demo (generated with matplotlib), not a Gazebo simulation or real hardware recording. It illustrates the target API and navigation behavior.

Try it now — 30 seconds, zero dependencies beyond numpy

git clone https://github.com/telleroutlook/3we-robot-platform.git
cd 3we-robot-platform && pip install -e sdk/threewe/
python examples/navigate_office.py

The recording was made with demo/record_demo.sh — clone the repo and run it yourself to verify.

---

5 Lines to a Moving Robot

from threewe import Robot

async with Robot(backend="mock") as robot:
    image = robot.get_image()
    await robot.move_to(x=2.0, y=1.0)
    pose = robot.get_pose()

Switch backend="mock" to backend="gazebo" or backend="real" — the Python API stays the same. Sensor models and sim configs may still need task-specific tuning; see Sim-to-Real Gap.

---

Why 3we?

If you are...	3we gives you...	Start here
AI/ML Researcher	Gymnasium envs, VLM/VLA integration, trajectory recording — focus on your model, not ROS2	Getting Started (AI)
Robotics Student	Full stack from PCB to Python, <$500 hardware, production-grade code instead of toy examples	Getting Started (Basic)
RL Practitioner	gymnasium.make("3we/Navigation-v1") — standard RL interface with sim-to-real presets for common tasks	Getting Started (AI)
Hardware Builder	Open BOM, assembly guide, CERN-OHL-P licensed PCB + structure	Assembly Guide

---

Quick Start

Status: The threewe SDK is under active development. pip install threewe is not yet available on PyPI. Install from source to try it today:

git clone https://github.com/telleroutlook/3we-robot-platform.git
cd 3we-robot-platform
pip install -e sdk/threewe/

import asyncio
from threewe import Robot

async def main():
    # "mock" backend works immediately — no ROS2 or Gazebo required
    async with Robot(backend="mock") as robot:
        # Navigate
        result = await robot.move_to(x=2.0, y=1.0)
        print(f"Reached: {result.success}")

        # VLM-powered instruction (requires: pip install threewe[ai])
        result = await robot.execute_instruction("go to the red door")

        # Get sensor data
        scan = robot.get_lidar_scan()
        imu = robot.get_imu()

asyncio.run(main())

Backends: Use backend="mock" to try the API instantly, "gazebo" for physics simulation (requires ROS2), "isaac_sim" for GPU-accelerated parallel simulation, or "real" for physical hardware.

Actual output (from the recording above)

$ pip install -e sdk/threewe/
Successfully installed threewe-0.1.0a0

$ python examples/navigate_office.py
============================================================
  3we Robot Platform — Office Navigation Demo
============================================================

[3we] Robot initialized  backend=mock  scene=office_v2

  Start pose: (1.0, 1.0)
  LiDAR: 360 rays, nearest obstacle: 0.99m

  Following path: 5 waypoints

[3we] Navigating to (2.0, 3.0)...
[3we] Planning path... distance=2.2m
[3we]   pos=(1.5, 2.1)  heading=63°  remaining=1.0m
[3we] Goal reached  distance=2.2m  time=4.5s
[3we] Navigating to (7.0, 3.5)...
[3we] Planning path... distance=5.0m
[3we]   pos=(4.0, 3.2)  heading=6°  remaining=3.0m
[3we] Goal reached  distance=5.0m  time=10.0s
...

  Final pose: (12.0, 10.0)
  Total distance: 20.5m
  Result: success

============================================================

RL Training

import gymnasium
import threewe.gym  # registers environments

env = gymnasium.make("3we/Navigation-v1")
obs, info = env.reset()

for _ in range(1000):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()

Benchmarks

threewe benchmark run --task pointnav --episodes 100 --backend gazebo

See the Benchmark Leaderboard for baseline results and submission instructions.

---

Capabilities

Capability	Status	Details
Consistent Python API across backends	✅ Ready	4 backends (mock/gazebo/isaac_sim/real), same Robot class — task-specific sim tuning still required (details)
Imitation Learning Data	✅ Ready	save_lerobot() export + HuggingFace Hub push/pull
VLM/LLM Control	✅ Ready	GPT-4o / Qwen-VL, async API decoupled from 50Hz control loop
Hardware Safety	✅ Ready	3-tier watchdog: 500ms cmd_vel timeout, 1s software WDT, 1.6s hardware WDT (TPS3813)
RL Gymnasium Envs	✅ Ready	4 standard envs (PointNav, Exploration, ObjectNav, VLN) + multi-agent
Edge AI Inference	✅ Ready	Hailo-8L M.2 accelerator (13 TOPS) on Raspberry Pi 5
VLA Model Deploy	✅ Ready	ONNX / PyTorch / Hailo HEF, from_pretrained() from Hub
Domain Randomization	✅ Ready	Physics, visual, and sensor noise — configurable per-episode
RGB-D Depth Camera	✅ Ready	get_rgbd_image() API, native depth in Isaac Sim
Modular Install	✅ Ready	threewe / threewe[sim] / threewe[ai] / threewe[all]

---

Architecture

flowchart TB
    A["<b>Your Python Code</b><br/>robot.move_to() · robot.get_image() · robot.execute_instruction()"]
    B["<b>threewe Python API</b><br/>Robot · Types · Config · AI (VLM/VLA) · Gym · Data · Benchmark"]
    C1["<b>MockBackend</b><br/>Zero-dep 2D<br/>kinematic sim"]
    C2["<b>GazeboBackend</b><br/>Gazebo Harmonic<br/>+ ros_gz_bridge"]
    C3["<b>RealBackend</b><br/>ROS2 Topics"]
    C4["<b>IsaacSimBackend</b><br/>GPU-accelerated<br/>parallel RL"]
    D["<b>ROS2 Jazzy + Nav2</b><br/>/cmd_vel · /scan · /odom · /camera · NavigateToPose Action"]
    E["<b>Firmware (ESP32-S3)</b><br/>Motor PID · Encoders · IMU · Safety Relay · micro-ROS"]
    F["<b>Hardware Layer</b><br/>Mecanum · DRV8833 · HC-SR04/LD06 · BNO055 · Battery · E-Stop"]

    A --> B
    B --> C1 & C2 & C3 & C4
    C1 & C2 & C3 & C4 --> D
    D --> E
    E --> F

Loading

---

Platform Comparison

Different tools for different needs. This table highlights where 3we fits — it is not a claim of superiority.

Feature	3we	TurtleBot 4	LeRobot	Isaac Lab
Python API (no ROS2 knowledge needed)	Yes	No	N/A	Partial
Same API, mock → sim → real	Yes	No	No	Yes
Open Hardware (PCB + BOM)	Full	Partial	N/A	N/A
Gymnasium Interface	Yes	No	Partial	Yes
VLM/VLA Integration	Yes	No	Yes	No
Hardware Cost	<$500	~$1200	~$2000+	N/A
Payload Hot-plug Bus	PBC-34	USB	N/A	N/A
Safety (HW E-stop)	Yes	Software	N/A	N/A
Community / Maturity	Early stage	Established	Active	Active

---

Hardware Specs

Spec	Standard v2
Drive	4-wheel Mecanum (65mm), omnidirectional
MCU	ESP32-S3 (dual-core 240MHz)
SBC	Raspberry Pi 5 (8GB)
AI Accelerator	Hailo-8L (13 TOPS)
LiDAR	LD06 (360°, 12m range)
IMU	BNO055 (9-axis fused)
Max Velocity	0.5 m/s linear, 1.0 rad/s angular
Battery	7.4V Li-ion, ~2h runtime
Safety	ISO 13850-inspired hardware E-stop + dual-channel relay (not certified)
Connectivity	Wi-Fi + BLE (+ 4G/5G optional)
Reproduction Cost	<$500

---

Repository Structure

3we-robot-platform/
├── sdk/threewe/          # AI-First Python package (pip install threewe)
│   ├── src/threewe/      #   Robot, Types, Backends, Gym, AI, Data, Benchmark
│   └── tests/            #   309 unit tests
├── examples/             # Ready-to-run demo scripts
├── firmware/             # ESP32-S3 firmware (ESP-IDF + micro-ROS)
├── ros2_ws/              # ROS2 packages (Nav2, SLAM, Gazebo, Perception)
├── hardware/             # Open hardware (KiCad PCB, BOM, mechanical)
├── sdk/payload_interface/# Payload communication library
├── sdk/web_control/      # TypeScript Web Components control panel
├── monitoring/           # Prometheus + Grafana observability
└── docs/                 # Guides, API reference, tutorials

---

Installation Options

Current: Install from source with pip install -e sdk/threewe/. The commands below will work once the package is published to PyPI.

# Core (types, Robot class, config)
pip install threewe

# With simulation (adds gymnasium)
pip install threewe[sim]

# With AI integration (adds openai, Pillow)
pip install threewe[ai]

# With data recording (adds h5py)
pip install threewe[data]

# Everything
pip install threewe[all]

For the real hardware backend, you also need ROS2 Jazzy:

# Ubuntu 24.04
sudo apt install ros-jazzy-desktop

---

Examples

Script	Description	Backend
navigate_office.py	Multi-waypoint navigation with verbose output	mock
hello_world.py	Connect, capture image, navigate	mock
rl_obstacle_avoidance.py	PPO training in simulation	mock
slam_exploration.py	Autonomous SLAM exploration	mock
data_collection.py	Record trajectories for imitation learning	mock
vlm_navigation.py	GPT-4o visual navigation	mock + API key
sim2real_demo.py	Same code, different backends	gazebo

Jupyter Notebooks

Step-by-step tutorials in notebooks/. Most require ROS2 + Gazebo unless noted:

Notebook	Topic	Requires
01_hello_world	Connect, sensors, basic navigation	Gazebo
02_slam_exploration	Autonomous mapping	Gazebo
03_point_navigation	Waypoints and path following	Gazebo
04_rl_training	Gymnasium + PPO training	mock (works standalone)
05_data_collection	Record trajectories, export to LeRobot	Gazebo

---

Roadmap

• Phase 1: ESP32 firmware + ROS2 stack + Hardware design
• Phase 1: threewe Python API + Mock backend
• Phase 1: Gymnasium environments + VLM/VLA integration
• Phase 1: Benchmark framework + Example scripts
• Phase 1: Gazebo/Isaac Sim backends (interface done, integration testing in progress)
• Phase 1: PyPI publishing (pip install threewe)
• Phase 2: On-robot status display panel (1.3" OLED SH1106, 3-key navigation, 5 pages, fault codes)
• Phase 2: Hardware Abstraction Layer for 3rd-party robots
• Phase 2: Foundation model fine-tuning pipelines
• Phase 3: 3we Hub (model/dataset sharing)
• Phase 3: Multi-robot fleet management

---

Blog & Dev Notes

Engineering deep-dives at 3we.org/blog:

Post	Topic
Dev Log #1: Why These Parts	ESP32-S3 vs STM32, mecanum wheels, PBC-34 bus design, honest project status
Sim2Real: Same Code to Real Hardware	Backend abstraction, domain randomization, validation protocol
Build a $300 ROS2 Research Robot	Complete BOM, assembly, software setup
30 Lines: Let GPT-4o Control a Robot	VLM perception-action loop, local model support

---

Licensing

Component	License	File
Firmware & Software	Apache License 2.0	LICENSE
Hardware Designs	CERN-OHL-P v2	LICENSE-HARDWARE
Documentation	CC BY-SA 4.0	LICENSE-DOCS

Attribution

Commercial use is welcome and free of charge. The licenses require that derivative works include attribution to this project. Recommended format:

This product is based on the open-source 3we Robot Platform.

For academic citations, see CITATION.cff.

---

Contributing

We welcome contributions! See CONTRIBUTING.md for development setup, branch strategy, and PR process.

---

Safety Notice

Warning

This platform contains moving mechanical parts and lithium batteries.

• Safety relay design is inspired by ISO 13850 but is not certified — do not use for safety-critical applications without independent certification
• Always verify E-stop function before operation
• Never bypass or modify the safety relay circuit
• Follow battery handling guidelines in documentation

---

Copyright 2025-2026 telleroutlook · 3we-robot-platform