Describe the skill. Shape the world. Keep the policy.
From a first attempt to a robot that knows what to do.
Duck + a first trick. Your selection stays here until you open Studio.
Tell alia what your robot should learn. It turns your idea into a task, builds a world and checks the setup. You follow the practice. You keep the policy.
See how it works in detail ↗Say what it should learn.
Pick a robot and describe the trick in a sentence. alia reads the skill, the objects and the goal out of your words.
See the plan. Shape the world.
Reward, success and failure in plain words, and a MuJoCo world with exactly the objects your task needs.
Watch an attempt become a skill.
PPO on our servers. Every checkpoint replayed in 3D, with a readable log and the learning curve.
Keep what works. Take it with you.
Test the policy in your browser, read an AI review of the run, and export ONNX for the real robot.
A ball, a goal, a flight of steps — or a whole warehouse aisle. Your words become physics objects. The task knows where they are. The robot gets room to figure them out.
Add the object. Move it a little. Try again. alia updates the scene and the task together, so the robot can see what you just changed.
The checks make sure objects start clear of the robot. You get to focus on the experience.
Kick, dribble, push
Push, move, carry
Climb up. Come back down.
Find your footing on an incline
Stop, avoid, go around
Give every attempt a destination
A point worth reaching
Turn a valve, screw or dial
Press a switch or doorbell
A belt that carries things away
Drop the ball through the ring
Racks, aisles, things to fetch
A surface to work on
Push it open, walk through
Slalom, cones and markers
Compose a new challenge from the world around it.
Concept scenes built from World Designer kits. In Studio, alia-built MuJoCo worlds carry through training, recorded video and the 3D viewer.
MuJoCo models in. Gymnasium tasks in the middle. ONNX policies out. Open formats all the way through, so your work can go wherever your robot does.
Read the docs ↗The same physics foundation, from server-side training to browser replay.
MJCF · URDF importA standard task API. Editable Python you can inspect, adapt and take with you.
OBSERVATIONS · ACTIONS · REWARDSRun inference on a Pi, a Jetson or a laptop. No training framework on the robot.
PORTABLE INFERENCE# On the robot — no torch, no SB3
# pip install onnxruntime numpy
import onnxruntime as ort
import numpy as np
session = ort.InferenceSession(
"walk_policy.onnx"
)
# Build obs with your exported task's schema.
obs = np.asarray(obs, dtype=np.float32)
action = session.run(
None, {"obs": obs[None]}
)[0][0]
send_joint_commands(action) # ~50 HzIntegration sketch: observations, action scaling and joint commands must match your task and robot.
Every robot starts somewhere. Drop this little one into a world with gravity, and see where it lands.
Interactive browser physics demo. No learned policy is applied. Training and MuJoCo runs live in the Studio.