FocalPolicy: Frequency-Optimized Chunking and Locally Anchored Flow Matching for Coherent Visuomotor Policy

Visuomotor policies aim to learn complex manipulation tasks from expert demonstrations. However, generating smooth and coherent trajectories remains challenging, as it requires balancing proximal precision with distal foresight. Existing approaches typically focus on optimizing intra-chunk action distributions, often neglecting the inter-chunk coherence. Consequently, inter-chunk discontinuities significantly impede the learning of coherent long-horizon actions. To overcome this limitation and achieve a synergetic balance between precision and foresight, we propose FocalPolicy, a foresight-aware visuomotor policy that combines Frequency-Optimized Chunking with Locally Anchored flow matching. We introduce a foresight composite objective that supervises time-domain alignment within the proximal actions while regularizing frequency-domain structure over multiple future action chunks to improve cross-chunk coherence. To efficiently learn complex action distributions, we design locally anchored sampling to enhance target signal propagation efficiency during consistency flow matching training. Extensive experiments demonstrate that FocalPolicy outperforms existing approaches and confirm the generalizability of our modules to other baselines. Project website: https://focalpolicy.github.io/

Teaching robots to perform long, complex tasks is like teaching someone a dance routine. Currently, robots often learn these movements in short "chunks." However, just as a beginner might stumble when transitioning between dance moves, robots struggle to connect these action chunks smoothly, resulting in jerky or failed motions during long-horizon tasks. In this paper, we introduce FocalPolicy, a new method that helps robots maintain both immediate precision and long-term foresight. Think of it as giving the robot two perspectives at once: a "zoom lens" to ensure the immediate next movement is exact, and a "wide-angle lens" to ensure the overall rhythm of future movements connects seamlessly. We also introduce a streamlined training technique that helps the robot learn these continuous motions much faster from expert demonstrations. Our experiments show that FocalPolicy allows robots to perform complex, multi-stage tasks—like pouring water or stacking objects—much more smoothly and reliably than previous methods. This work brings us closer to creating dependable robotic assistants for real-world applications.