FlowDPG: Deterministic Policy Gradient on Flow Matching Policies for Real-World Manipulation
The episode discusses FlowDPG, a method for deterministic policy gradient on flow matching policies designed to solve computational and numerical fragility caused by backpropagating through time. Hosts explore how FlowDPG distills critic gradients into a velocity field to ensure stable updates, using dual-correction mechanisms for feasibility and value correction. The paper shows superior performance in complex manipulation tasks.