Robofauna
Walk
One policy handles forward, backward and turning on joystick commands. In simulation it walks forward at 0.14–0.17 m/s.
Walking in sim
Open-source physical AI
Learn in simulation.
Learn in simulation.
Move in the real world.
Robofauna builds open-source physical AI robots. Train a policy with reinforcement learning in simulation, run it on a real NanoEmu, and follow the arm as it grows toward a humanoid. CAD, firmware, BOM and training code are all public.
01 · Physical AI
NanoEmu
A 25 cm biped on 14 bus servos. Its walking policy is trained with reinforcement learning in simulation, built to run on the robot itself.
Meet NanoEmu 02 · LearningRL Lab
Reinforcement learning in the browser. MuJoCo physics and PPO run on your own GPU, with nothing to install.
See RL Lab 03 · ManipulationFauna Arm V1
A human-scale robot arm with an adult's 700 mm reach and a 4 kg hold, for about the price of a laptop.
See the arm
NanoEmu · Physical AI biped
A biped that learns in simulation, then walks for real.
NanoEmu is a 25 cm, 14-servo biped with reverse-knee legs and a beak that opens and closes. Its walking policy is trained with PPO reinforcement learning in MuJoCo, then made to run on the robot's own ESP32-S3 at 50 Hz with IMU feedback. Closing that simulation-to-real gap is what we are building, and the CAD, firmware and training code are all open.
$299
kit · assembled $369 · pre-order
≈ 250 mm
Concept render · final shell in CAD
≈250mmStanding height
14SCS0009 bus servos
50HzRL policy, trained in simulation
ESP32‑S3Controller, camera, Wi‑Fi
$299Kit · assembled $369
Sim-to-real
One policy, from simulation to a real body.
- 1Learn in simulation
PPO reinforcement learning in MuJoCo, thousands of robots at once in RL Lab.
Walking in sim - 2Run on the real robot
The trained policy runs on NanoEmu's own ESP32-S3 at 50 Hz, steered by IMU feedback.
Sim-to-real in progress - 3Keep adapting
The Base adds slow, on-device learning of behaviour once the robot is home.
Concept
Designed to print
Rounded shells, few parts and exposed joints, all sized for a hobby 3D printer.
What it does
Walks today. Learning the rest.
We mark each skill by where it actually stands. Walking comes from a PPO policy trained in MuJoCo, and the rest are on the way.
Balance
Holds its stance after a 5 cm drop or a push, instead of just standing still in a calm pose.
In trainingPick up
Bends down and grabs a sock off the floor with its beak. The pose reaches the ground in simulation.
In developmentGet back up
Righting itself after a fall needs more torque and joint range than today's servos give. We are working on it in the open.
Research
Coming soon · NanoEmu Base
A pet that grows up with you.
NanoEmu can't talk. It only acts. Add a NanoEmu Base, a small AI computer that lives next to it, and it starts growing the way an animal does: call its name often enough and one day it answers to it.
Concept stage · specs, price and date not set
- Learns from living with you.A camera and microphone, and later touch sensors, collect the day. While NanoEmu sleeps on its charger, the Base learns a little from it, every night.
- No goals, no training menu.Growth follows what happens, not a target. It may pick up odd habits and make mistakes, like a real animal. You can reset it, but you can't steer it.
- Behavior only.Walking, sitting and playing football are separate motor policies you train in RL Lab, and they run on NanoEmu's own ESP32. The Base only sends a behavior style, and the ESP32 blends its policies into movement. Growth shapes what NanoEmu chooses to do, not how its legs move.
- A nest that charges.The Base is an egg about 21 cm tall, roughly three times the volume of a typical mini PC. It sits in a nest that also charges NanoEmu, and that is where the nightly learning happens.
- Private and offline.Everything stays on the Base, and no internet connection is needed. The plan: encode what it sees and hears into compact learning data, delete the original video and audio right away, and delete that data once it has learned from it.
Inside the kit
Built from parts you can source yourself.
Every part is off the shelf or 3D-printed. Buy the kit to skip the sourcing, or order parts one by one from the shop and print the shell yourself.
- Actuators14 × Feetech SCS0009 serial bus servos
- LegsHip roll → yaw → pitch, reverse knee, ankle. Hip roll runs through a 2:1 linkage for twice the torque.
- HeadNeck pitch, head pitch, head yaw and a beak gripper
- ControllerSeeed XIAO ESP32-S3 Sense with camera, Wi-Fi and Bluetooth
- SensingBMI270 IMU for angular rate and gravity direction, plus foot contact
- Power2S LiPo (8.4 V) with a 6 V buck converter for the servos
- ShellAbout 180 g of PLA or PETG in blocky voxel styling: ivory, black and yellow
Open BOM
Where the money goes
Servos are about 69% of the parts cost. We publish the BOM so you can check our price, or skip the kit and build from scratch.
Servos $190
Controller, IMU, driver $41
Battery, buck $17
PCB, print, hardware ≈$28
Build it yourself
Retail parts, one unit
Buying every part at retail today for a single robot costs about this much. The kit price is close to it, and every design file is free.
≈$276 retail BOM · kit $299
RL Lab · Browser reinforcement learning
Teach your robot new tricks, from a browser tab.
RL Lab trains walking and balance policies on your own CPU or GPU, with nothing to install and no cloud. Each robot gets its own experiment, with one policy per goal: stand without falling, or walk on command.
- Zero install, zero cloud. Everything runs on your own CPU or GPU, and robot files never leave your device.
- Two physics engines. MuJoCo for reference accuracy, or experimental WebGPU physics that trains thousands of robots at once.
- Watch it learn. The latest policy drives the robot in 3D while it trains. Steer it with an 8-way joystick or WASD, push it, drop it.
- Honest metrics. Reward curve plus real speed, command error, fall count and a walking check.
- Servo editor. Set zero, direction, limits and start pose for each joint to match your real servos.
- Sim to real, next. Export a trained policy to the ESP32 in your NanoEmu.
Development build screens · rl.robofauna.com
4,096Robots training at once
1–2hTo a walking policy
17–29kPhysics steps per second
0Installs or cloud GPUs
LocalYour files stay on your device
A · now
NanoEmu policies
Walking on command works and passes an independent MuJoCo check. Sit and stand are next.
B
ESP32 sim-to-real
The same policies run on a real NanoEmu.
C
Your robot
Import your own MuJoCo XML, from a robot dog to a hexapod or humanoid, and design rewards by hand or by chatting with AI.
D
Policy sharing
Upload, download and remix robots and policies with the community.
E
Robot arms
Fixed-base reaching and tracking tasks.
Fauna Arm V1 · Human-scale manipulator
An adult's reach, at the price of a laptop.
From shoulder to fingertip Fauna Arm V1 spans about 700 mm, close to the arm of a 175 cm adult. It holds 4 kg fully extended. It is open hardware, and the full kit is $1,290.
Concept render
$1,290Full kit
700mmShoulder to fingertip
4kgHold, fully extended
270 / 230Upper / forearm, mm
OpenCAD, firmware, BOM
Open source
Everything is on GitHub.
Models, training code, design notes and the reasons behind each decision. Fork them, print them, improve them.
Roadmap
One workshop, growing.
Each robot reuses the servos, boards and training stack of the last, so every project makes the next one cheaper.
- NowIn progress
NanoEmu walking policies
GPU training walks in simulation. Next: firmware on the ESP32 and the first steps on real hardware.
- NextPre-order
NanoEmu kit ships
Final shell CAD, leg PCB and BOM locked, then pre-orders are fulfilled.
- NextDesigning
Fauna Arm V1
A human-scale robot arm kit.
- LaterConcept
NanoEmu Base
A small AI computer that lets NanoEmu grow habits of its own, learning a little every night while it charges.
- LaterPlanned
RL Lab for any robot
Bring your own robot model, design rewards in plain language, and share policies.
- LaterPlanned
Humanoid
Joints, boards and gearboxes from every earlier robot come together.
FAQ
Questions
Does NanoEmu come assembled?
Your choice. The kit ($299) is DIY: you assemble the legs, head and body, wire the servo bus and flash the ESP32, with a guide published alongside the CAD. The assembled version ($369) arrives built, with all 14 servos calibrated, firmware installed and joints checked.
What does "physical AI" mean for NanoEmu?
NanoEmu's movement is not scripted. It is a policy learned with reinforcement learning in a physics simulator, then run on a real body with servos and an IMU. Bridging the difference between simulation and the real robot (sim-to-real) is the core of the project. Everything is open, so you can retrain and change it yourself.
Do I need a 3D printer?
Not if you buy the printed shell set in the shop. If you have a printer, print the shell yourself in PLA or PETG (about 180 g) and buy only the parts you need.
Can it really walk?
It walks in simulation today with a PPO policy trained in MuJoCo, forward, backward and turning. Running that policy on the real robot (sim-to-real) is the current step. We will post progress openly.
Do I need a powerful computer to train it?
RL Lab runs in a browser and uses whatever GPU you have through WebGPU. More GPU means more parallel robots and faster training, but no cloud and no install are needed.
What's the license?
CAD, firmware, BOM and training code are published on GitHub. The license for each repository will be listed there before the kits ship.
What is the NanoEmu Base?
A separate, coming-soon add-on: a small AI computer that sits next to NanoEmu, watches and listens through its camera and microphone, and learns a little every night while the robot charges. NanoEmu stays a robot that acts rather than speaks, and it grows its own habits, including responding to its name. Everything stays on the Base and it works without internet. The plan is to turn what it sees and hears into compact learning data, delete the original video and audio right away, and delete that data once it has learned from it. Walking and movement policies keep running on NanoEmu's own ESP32. It is still at the concept stage, so details may change.
When will pre-orders ship?
We will announce a shipping window once the final shell CAD and leg PCB are locked. Sign up through the shop to hear first.
Build an emu. Teach it to walk.
Kit $299 or assembled $369, pre-order. Parts are also sold one by one.