Pick a ready-made project on your dashboard — Go2 PPO, G1 AMP, and more — or describe the robot and the behaviour you want and let AI Build assemble it. Both paths land on the same canvas, and every value stays yours to edit.

Note: the training backend is chosen when the project is created and cannot be changed later. Everything else can.

Open the Robot node to pick your machine, or bring your own: upload a model, or import one from a GitHub link. Review the joints, the PD response and the starting pose before you spend a single GPU-hour on them.

Note: the model itself — links, joints, physical parameters — is editable online too, in the robot editor.

The Environment node is the ground your robot learns on: flat, rough or stairs, with its size, its friction, and how many copies train in parallel. Sculpt your own field once and reuse it in every project.

Double-click Task to step inside it — a canvas within the canvas, same camera, same nodes. This is where you say what the robot should learn:

• Movements — Continuous for what it does all the time (walk, turn); Triggered for what it fires on demand (jump, crouch). A triggered move is a skill; you never declare it twice.
• Rewards — a green block for what the policy is paid for, a red one for what it is charged for. Wire a set to the movements it applies to.
• Terminations — when an episode ends, for the whole canvas or for one movement.

Back on the main canvas, the Observation node decides what the policy is allowed to sense — the sensors, and nothing beyond them.

The Trainer node holds the learning rate, the iteration count, the network shape and the rest. Start from the template values and move one at a time. The algorithm is derived, not switched: attach a reference motion clip in Task and the run becomes AMP — nothing to remember, and no way to get the two out of step.

Press Run and the Start training dialog asks what only now can be answered: what to call this bundle — an orange tag warns you when the name overwrites one you already have — and where the run goes: our GPUs, your own server, or your own machine. Advance holds the content switches.

Note: a machine that cannot take this run says so here, and points at the one setting that would fix it. Cancel changes nothing.

Confirm, and it runs. Reward, episode length and losses stream into the charts as they happen, so a bad run costs you minutes instead of a night. Read the curves, change one thing, run it again.

When the shape looks right, replay the trained policy in your browser before it ever touches hardware.