First robot experiment

Use this limited local workflow to move from a robot definition to an evaluated joint-feedback controller. You need Studio, an installed MuJoCo runtime, and a robot URDF with its referenced meshes and suitable physical properties.

Build and validate

1. Create a project.
2. Open Builder, import the robot and meshes, and inspect links and joints.
3. In Configure, supply missing mass, inertia, joint limits, and collision information. In Test, resolve validation findings.
4. Save a robot version. Joint sliders preview kinematics; they do not run physics.

Simulate

1. Open the saved version in Simulation.
2. Set duration, gravity, and joint position targets, then run.
3. Replay the result and inspect the motion. A completed simulation does not establish task success.

Train and evaluate

1. Open Experiences → Loops → Training experiments.
2. Create an Experience and Loop, then save an environment and a joint-reaching task.
3. Save an experiment with the Builder version and training configuration.
4. Choose Local · MuJoCo, launch training, and inspect the evaluation, baseline comparison, and replay.
5. From a completed run, choose Manage policy in Deployments to preserve the candidate and its evidence.

The current CEM trainer optimizes a joint-feedback controller for a fixed-base robot with 1–32 finite revolute/prismatic joints. It does not consume demonstration datasets or produce an image-based policy. Training does not guarantee improvement.

Registering or assigning the policy does not execute it on physical hardware. Continue with Experiences and Deployments for the workflow boundaries.