single source · training objective
Multiple flow-time samples per action chunk
A training-objective technique for flow-matching action experts that evaluates the flow loss at several sampled noise levels for each action chunk instead of one.1,2 Multiple (noise, target) pairs are drawn per example so the same action chunk and visual-language context supervise several points along the flow trajectory.3
Trade-off More flow samples raise average performance but cost GPU memory, which is why the fine-tuning stage used fewer samples than post-training.4,5
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
For each robot action chunk, we evaluate the flow objective at multiple noise levels
— MolmoAct2 Action Reasoning Models for Real-World Deploymentvary the number of flow samples per action chunk
— MolmoAct2 Action Reasoning Models for Real-World Deploymenteach robot sample contributes four points on the same flow trajectory while reusing the same visual-language context
— MolmoAct2 Action Reasoning Models for Real-World Deploymentincreasing generally improves average performance
— MolmoAct2 Action Reasoning Models for Real-World Deploymentmainly due to GPU memory constraints
— MolmoAct2 Action Reasoning Models for Real-World Deployment
Also called multiple flow samples.
Filed under training objective