Event

FlowVLA improves robot quantitative pouring accuracy with physical completion estimation

1 reports · 1 sources

The FlowVLA paper author team proposes the FlowVLA method, which separates motion generation from physical completion estimation to address the stopping decision problem in robot quantitative liquid pouring. The method uses a shared π0.5 action expert to learn approach, pour, and return actions, and uses a goal-agnostic estimator that fuses vision, joint, wrist torque, and recurrent force belief history to estimate the transferred mass, triggering the return when the estimated value plus the predicted residual outflow reaches the requested mass. In the coffee-pot/water experiment, FlowVLA achieved a target-balance MacroMAE of 13.1 g, lower than all tested VLA baselines; in all attempts, 53.7% both completed the task and kept the mass within ±10% of the requested mass. The same model can interpolate to unseen 150/250 g targets without retraining, and another independently trained model evaluated the method in three additional container-liquid settings.