Invertible Neural Network Adapter for One-Step Flow Matching in Robot Manipulation

Method

Following [ 13 ] , the forward process of the flow ODE is defined as a straight-line path between the data distribution and the noise distribution.

Specifically, given a data sample $\mathbf{x}_{1}\sim\mathcal{D}$, a noise sample $\mathbf{x}_{0}\sim\mathcal{N}(\mathbf{0},\mathbf{I})$, and a time step $t\sim\mathcal{U}[0,1]$, the interpolated state $\mathbf{x}_{t}$ is defined as The corresponding velocity is defined as the displacement from the noise sample to the data sample: The flow model parameterized by $\theta$ is trained to predict the ve...

Experimental Results

Five dexterous manipulation tasks from RoboTwin 1.0 [ 20 ] are selected to evaluate the manipulation capabilities of the proposed adapter. For 2D image observations, comparisons are conducted against ManiFlow, Diffusion Policy, and Flow Matching Policy, all built upon a shared ResNet-18 encoder to ensure a fair comparison.

For 3D point cloud observations, the proposed method is further compared with ManiFlow, 3D Diffusion Policy, and 3D Flow Matching.

Limitation

While the invertible neural network adapter improves performance across both simulated and real-world manipulation tasks, several limitations remain. First, the method assumes that the action distribution can be mapped to a well-structured latent space via an invertible transformation.

Although this holds for the manipulation tasks studied here, its validity for highly multimodal or discontinuous action distributions remains an open question.

Sources and Demonstrations

Figures are reproduced from Zhang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.