Toward Agile Maneuvers in Highly Constrained Spaces: Learning From Hallucination

While classical approaches to autonomous robot navigation currently enable operation in certain environments, they break down tightly constrained spaces, e.g., where the needs engage agile maneuvers squeeze between obstacles. Recent machine learning techniques have potential address this shortcoming, but existing require vast amounts of experience for training, during which must operate close proximity obstacles and risk collision. In letter, we propose side-step requirement by introducing a new paradigm called from hallucination (LfH), can use training data collected completely safe environments compute controllers that result fast, smooth, highly environments. Our experimental results show proposed LfH system outperforms three baselines on real generalizes well unseen including those based both techniques.