Once you have a dataset version you trust, training is a few clicks away. You pick the version and a policy, EmbodiFlow packages the data and runs the job on a GPU, and you watch the loss curve, logs and checkpoints fill in on the model page.
Training only ever runs on a frozen dataset version. The data can't change under a running job, and every model remembers exactly which version it was trained on, so "what data made this model?" always has an answer.
Start a training run
Open the dataset version
In your project, open Versions and choose the version you want to train on. In the top-right corner, next to Export, choose Train model. You can also start from Training in the sidebar: choose New training, then pick the project and the version there. Both lead to the same settings.
Choose a policy
The Policy cards list what the form offers, each with the GPU memory it needs and whether your episodes need task instructions. ACT is marked Recommended: it learns short action chunks from demonstrations, and it's the policy we recommend starting with. The form also offers SmolVLA, π0 LoRA, π0.5 LoRA, GR00T and WALL-OSS, and shows Diffusion as Not available yet. Which of these can actually run depends on the training images installed for your team, so check with your administrator before you pick one other than ACT.
Pick a GPU and a length
Under GPU, choose a card that meets the policy's minimum. Under Length, Standard is the default. Quick becomes available after a first measured run of that policy on that GPU, because until then nobody knows how long a step takes. For exact numbers, open Advanced and set Training steps, Batch size and Seed; leave the steps empty to use the length you picked.
Read the notices under the settings before you start. If your version has no test episodes, the dialog warns that the model can't be evaluated afterwards. Training still works, you just won't get an evaluation.
Start training
Choose Start training. If you started from the version page, the new model's page opens; if you started from Training, the training job page opens. The job waits for a GPU and then starts on its own. Waiting for a GPU is a normal state, not a failure.
Watch it train
The model page has two tabs, and Training is where the action is. At the top you see the status (Queued, Running, Succeeded, Failed and so on) with the latest step. Below it, the Metrics chart plots the loss against the step and updates by itself while the job runs. When the job ends, the page says "Training is no longer running. Live updates stopped." and shows the Final step, Final loss and number of Checkpoints.
A falling loss tells you the model is fitting your demonstrations. It doesn't tell you how the policy will behave on a robot, so treat it as a health check, not a score.
Further down, Checkpoints lists each saved checkpoint with its step, time and size, and Logs shows the training output. You can search the logs, page through them with Older and Newer, switch on Show raw logs for the unformatted text, or choose Download .txt to keep a copy.
While a job runs, Cancel training stops it. If a job failed after saving a checkpoint, Resume from last checkpoint carries on from there. To try other settings, Train again with changes opens the same dialog on the same version.
The model overview
The Overview tab sums the model up. Chips show the policy and the robot it was trained on (for example ACT and Trained on TRON2 AV3), followed by the final step, loss and checkpoint count. On the right you find its ID, the training job, the robot protocols it can be deployed over (Deploy via) and when it was created. Open training takes you back to the curve.
Download a checkpoint
Open Training in the sidebar and choose the job. Its page has Overview, Logs, Metrics and Checkpoints tabs. On Checkpoints, choose Download on a row to save that checkpoint. The same page lists every attempt, so a retried job keeps its failed attempt next to the new one.
From a terminal, the CLI streams a checkpoint to disk and verifies its size and checksum: embodiflow training checkpoints download --job <job-id> -o weights.tar.
When a job fails
Open the job from Training. The failed attempt gives the reason in plain words, with the code under Technical details. Fix the cause, then choose Retry. If the reason says the training image could not be pulled, the job never started: ask your administrator to check access to the image registry, then retry.