Tell it what your robot should learn.
Get the policy.

The robot model supplies the physics: bodies, joints, motors, masses. It stays fixed; everything alia writes wraps around it. Start with a tiny robot like the ones on a maker’s bench — Open Duck Mini, Stanford’s ToddlerBot, the LeRobot SO-101 arm, a Crazyflie — pick one of seven fast-training classics, or import your own MJCF / URDF. Bigger robots such as Spot, G1 and Panda are there as previews: simulation only, with short capped runs, until GPU training arrives.

Your own robot: Robots → Import, then pick it like any built-in one.

From your brief, alia writes a Gymnasium task in Python around your robot: what the policy observes, what its actions do, the reward it earns every step, when an episode succeeds, and when it ends early. The Plan page explains each part in plain words — reward terms, success criteria, failure conditions, observation signals — and you change any of it by talking to alia.

class KickBallEnv(MjcfRobotEnv):
    ROBOT_ASSET = gym_asset("Walker2d")
    SCENE = [{"type": "ball", "name": "ball", "pos": [1.5, 0.0]},
             {"type": "goal", "name": "goal", "pos": [5.0, 0.0]}]

    def compute_reward(self, action):
        to_ball  = -self.object_distance("ball")
        ball_fwd = self.object_position("ball")[0] - self.object_initial_position("ball")[0]
        upright  = 1.0 if self.upright > 0.8 else 0.0
        return 0.5 * to_ball + 2.0 * ball_fwd + 0.1 * upright - 0.001 * (action ** 2).sum()

Illustrative — the real code is generated for your robot and brief, and is always visible on the Plan page.

The trained policy exports as a deterministic ONNX file with observation normalisation baked in, plus a metadata file with shapes and checksums. On the robot you only need onnxruntime and numpy — no PyTorch, no Stable-Baselines3.

pip install onnxruntime numpy

import onnxruntime as ort
sess = ort.InferenceSession("kick_ball_policy.onnx")
action = sess.run(None, {"obs": obs[None]})[0][0]
send_joint_commands(action)   # at the robot's control rate

Simulation is not reality. Start at low power, keep people clear and keep an emergency stop in reach.