Autonomous Learning Robots D3IL
A Benchmark for Imitation Learning from Multi-Modal Human Demonstrations
D3IL provides a variety of robotic manipulation simulation environments and a large number of high-quality human demonstrations for each. These demonstrations, are explicitly multi-modal, i.e., cover various distinct solutions for the task. Variations include different paths to a goal and a different order in which objects are sorted, among others. As such, they are particularly suited for the evaluation of e.g. diffusion policies.