New flow matching method boosts robot success from 0% to 70% on complex tasks

Flow matching policies for robot manipulation learn continuous velocity fields to transform noise into actions during inference, but standard training optimizes pointwise velocities—creating a mismatch when the same field is numerically integrated at test time. This leads to compounding trajectory errors, especially in long-horizon tasks. Ahmed et al. introduce four complementary fixes: auxiliary rectified flow velocity regression for uniform temporal supervision, multi-step trajectory consistency training that directly supervises integrated displacements, velocity field regularization for smoothness, and fourth-order Runge-Kutta (RK4) inference to reduce discretization error. The authors show that none of these components works alone—RK4 without a smooth field fails, and smoothness without trajectory-level supervision still drifts. The method also pairs a dual-view 3D point cloud encoder using two independent PointNet encoders for better spatial perception.

Tested on four real-robot tasks with a Franka arm and a Boston Dynamics Spot, the approach achieved 70% and 60% overall success on two long-horizon multi-phase tasks where both baselines scored 0%, and reached 100% on precision tool placement. Three MetaWorld simulation tasks confirmed consistent improvements. The work underscores that trajectory-level supervision is essential for reliable policy execution, offering a practical recipe for closing the train-inference gap in visuomotor learning.