Roboflow adds Cosmos 3 Edge training for video actions

The release describes an Action Recognition project: label videos, create a dataset version, then choose Cosmos 3 with Edge size for training. After completion, use the trained model’s identifier to submit a new clip through the Serverless Cloud API. Compare its predicted timeline with the prescribed procedure to identify omitted, repeated or misplaced steps.

The annotation guide explains the preparation. Each label combines an action class with a time interval. Its box-opening example starts when the lid moves and ends when the box is open, excluding nearby idle time. Drag across the timeline and assign the class, then play through the action to check both boundaries. For precise adjustments, position the playhead, right-click the segment, and select Set In for its start or Set Out for its end. Edits save automatically, but wait for Saved before leaving.

The training guide adds a budgeting constraint. Cosmos 3 Edge video-training estimates depend on clip count, duration and sampling settings; missing video details mean no estimate. A credit cap requires a usage-based billing plan. Reaching it stops training while retaining model weights. Each model has a minimum acceptable cap. Choosing a cap below the estimated cost is permitted, but the run will likely end early rather than complete the planned training.