Title: AffordGen: Creating Varied Demonstrations for Robust Object Manipulation via Affordance Alignment

Abstract: While contemporary imitation learning techniques have demonstrated significant achievements in robotic manipulation, their effectiveness is frequently hindered by geometric inconsistencies stemming from insufficient data variety. To address this challenge, the AffordGen framework harnesses the capabilities of advanced 3D generative models alongside vision foundation models (VFMs). It achieves this by exploiting semantic correspondences of key points across extensive 3D mesh datasets to synthesize novel manipulation trajectories for robots. This resulting large-scale dataset, characterized by its awareness of affordances, facilitates the training of a resilient closed-loop visuomotor policy. This approach effectively merges the semantic adaptability of affordances with the reactive stability inherent in end-to-end learning systems. Testing conducted in both simulated and real-world environments reveals that policies developed using AffordGen attain high success rates and facilitate zero-shot generalization to completely novel objects, thereby markedly enhancing the efficiency of data utilization in robotic learning. Project Page: https://jiaweiz9.github.io/AffordGen-release/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC