VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot Manipulation
VQActFlow is a multi-task robot policy that learns to select different manipulation modes by tokenizing continuous actions into discrete codes using a VQ-VAE. It uses Variational Flow Matching to generate action sequences while explicitly maintaining a preference over these modes. Inference guidance, specifically classifier-free guidance and a learned codebook critic, steers the policy toward the desired task action and ensures feasibility in the physical scene.