Rapid advances in perception have enabled large pre-trained models to be used out of the box for processing high-dimensional, noisy, and partial observations of the world into rich geometric representations (e.g., occupancy predictions). However, safe integration of these models onto robots remains challenging due to a lack of reliable performance in unfamiliar environments. In this work, we present a framework for rigorously quantifying the uncertainty of pre-trained perception models for occupancy prediction in order to provide end-to-end statistical safety assurances for navigation. We build on techniques from conformal prediction for producing a calibrated perception system that lightly processes the outputs of a pre-trained model while ensuring generalization to novel environments and robustness to distribution shifts in states when perceptual outputs are used in conjunction with a planner. The calibrated system can be used in combination with any safe planner to provide an end-to-end statistical assurance on safety in a new environment with a user-specified threshold $1-\epsilon$. We evaluate the resulting approach - which we refer to as Perceive with Confidence (PwC) - with experiments in simulation and on hardware where a quadruped robot navigates through indoor environments containing objects unseen during training or calibration. These experiments validate the safety assurances provided by PwC and demonstrate significant improvements in empirical safety rates compared to baselines.

翻译：感知技术的快速发展使大规模预训练模型能够直接用于处理高维、嘈杂且不完整的观测数据，将其转化为丰富的几何表示（例如占据预测）。然而，由于这些模型在陌生环境中缺乏可靠性能，将其安全集成到机器人上仍然充满挑战。本文提出一个框架，用于严格量化预训练感知模型在占据预测中的不确定性，从而为导航提供端到端的统计安全保障。我们基于共形预测技术构建校准感知系统，该系统轻量化处理预训练模型输出，同时确保对新环境的泛化能力，并在感知输出与规划器结合使用时对状态分布偏移具有鲁棒性。校准系统可与任意安全规划器结合使用，在新环境中提供以用户指定阈值$1-\epsilon$为界限的端到端统计安全保障。我们通过仿真实验和硬件实验（四足机器人导航穿过训练或校准中未见的室内物体环境）评估了所提出的方法（称为自信感知，PwC）。这些实验验证了PwC提供的安全保证，并表明与基线相比，经验安全率显著提升。