iGibson, a Simulation Environment for Interactive Tasks in Large Realistic Scenes.

摘要: We present iGibson, a novel simulation environment to develop robotic solutions for interactive tasks in large-scale realistic scenes. Our contains fifteen fully home-sized scenes populated with rigid and articulated objects. The are replicas of 3D scanned real-world homes, aligning the distribution objects layout that real world. iGibson integrates several key features facilitate study tasks: i) generation high-quality visual virtual sensor signals (RGB, depth, segmentation, LiDAR, flow, among others), ii) domain randomization change materials (both texture dynamics) and/or their shapes, iii) integrated sampling-based motion planners generate collision-free trajectories robot bases arms, iv) intuitive human-iGibson interface enables efficient collection human demonstrations. Through experiments, we show full interactivity agents learn useful representations accelerate training downstream manipulation tasks. also enable generalization navigation agents, imitation learning simple demonstrated behaviors. is open-sourced comprehensive examples documentation. For more information, visit our project website: this http URL