In this paper, we present an end-to-end planning framework based on a novel imperative learning (IL) approach. The method involves a bi-level optimization (BLO) process that combines network update and metric-based trajectory optimization during training to produce smooth and collision-free trajectories using only a single depth measurement. The IL is able to utilize task-level loss and optimize through direct gradient descent. This allows the method to be trained in an efficient unsupervised manner, eliminating the need for explicit trajectory labels.

Paper: https://arxiv.org/abs/2302.11434
Code: https://github.com/leggedrobotics/iPl...