Robofauna
Zero install, zero cloud
All computation runs on your device. Robot files are never sent to a server.
RL Lab · Browser reinforcement learning
Train in your browser. Move in the real world.
Open a tab and your own CPU or GPU trains a small biped to stand and walk with reinforcement learning (PPO), the simulation side of NanoEmu's sim-to-real pipeline. Watch every policy move in 3D as it learns. No Python setup, no cloud GPU, no robotics degree.
rl.robofauna.com · opens in a new tab
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
Who it's for
For people who build their own robots.
Makers, students and researchers who want a small ESP32-and-servo robot to walk. Most open biped projects train in Python on a dedicated GPU machine. RL Lab needs only a browser.
Makers
You built the robot. Now make it walk without setting up a training rig.
Students
See what reinforcement learning actually does, one policy update at a time.
Researchers
Change rewards and PPO settings, keep training, and compare policies side by side.
How it works
From an empty tab to a walking robot.
Each robot gets its own experiment, and each experiment holds one policy per goal.
- 1
Create an experiment
Pick a robot from the library. NanoEmu, a 14-joint biped, is ready today.
- 2
Choose a goal
Stand without falling, recovering from a 5 cm drop or a push. Or walk on command: forward, backward and turning.
- 3
Start training
Train with MuJoCo physics for reference accuracy, or experimental WebGPU physics that runs thousands of robots at once.
- 4
Watch and test
The latest policy drives the robot in 3D. Steer with an 8-way joystick or WASD, push it, drop it, and check the numbers.
- 5
Save and continue
Experiments save in the browser automatically. Export them to a file, change settings and keep training.
A look inside
The screens you train with.
Captured from the current development build. Names and the robot model will change before launch.
What's inside
Everything you need to train, in one tab.
Built with TypeScript, MuJoCo WASM, WebGPU (WGSL), TensorFlow.js, Three.js and IndexedDB.
Two physics engines
MuJoCo is the accurate reference. WebGPU physics is experimental and runs about 17,000–29,000 steps per second.
Live 3D view
The newest policy moves the robot while training continues, so you see progress as it happens.
Honest metrics
Reward curve, real speed, command error, fall count and a walking check, not just a rising number.
Servo editor
Set zero, direction, min and max angle and the start pose for each joint to match your real servos.
Experiments & policies
Keep several policies per robot. Everything restores on reload, and exports to a file.
Advanced settings
Learning rate and other PPO settings, per-term reward weights and walking command ranges.
Joystick & balance tests
Command walking in eight directions, or push and drop the robot to test balance recovery.
Result so far
A walking policy in one to two hours.
Training 4,096 NanoEmu robots at once on a GPU produced a policy that follows forward, backward and turning commands in one to two hours. It passes an independent MuJoCo check. Recovery from pushes is still being improved.
WebGPU physics needs a browser with WebGPU support. MuJoCo training runs on the CPU.
Roadmap
Where RL Lab is going.
Train in the browser, test in the browser, then run it on the real robot and share it.
A · now
More basic policies
Walking works. Sitting down and standing up are next.
B
ESP32 sim-to-real
Export a policy from the web as an ESP32 file and run it on the real robot.
C
Your robot
Import your own robot XML, from a dog to a hexapod or humanoid, and design rewards by hand or by chatting with AI.
D
Policy sharing
Upload robots and policies, download them and keep training someone else's (remix).
E
Robot arms
Fixed-base reaching and tracking tasks.
Ready to train?
RL Lab opens in a new tab. Everything runs on your machine.