Computer Vision News - December 2021

5 Rapid Exploration for Open-World Navigation Data The authors carefully approach the problem of using previous experience to build a robust method, by curating their own dataset made of over 5,000 self-supervised trajectories. This accounts to 9 hours, collected over 9 distinct real-world environments of varying complexity (see Figure below d) Training Environments, e) Unseen Test Environments). The dataset contains measurements from a wide range of sensors, including more and less accurate ones, following the assumption that learning-based techniques coupled with multimodal sensor fusion can provide a lot of benefits in the real world. This dataset is also made available online, contributing to open-source and replicable research. To gather these data, the authors use a time-correlated random walk and a mechanism to detect if the robot is in collision or stuck and then an automated backup manoeuvre that drives the robot out. The collision detectors are further used to generate event labels for the collected trajectories. Implementation The method introduced in this paper is called RECON ( R apid E xploration C ontroller for O utcome-driven N avigation). It is based on two novel elements: • An uncertainty-aware, context-conditioned representation of goals that can quickly adapt to novel scenes. This is also referred as the latent goal model, which encodes prior knowledge about perception, navigational affordances, and short-horizon control. • A topological map with memory of the target environment. These serve a two-fold purpose: exploring a novel environment which uses a combination of frontier-based exploration and latent goal sampling with the learned model, and navigation of an explored environment using the topological graph and the learned model. All of this is made possible through an initial step of supervised training using prior experience on previously visited environments provided by the acquired dataset. D E 6WDUW *RDO ,PDJH 0DS 1RGH ([SORUDWLRQ 3DWK 'LVFRYHUHG 3DWK F H 8QVHHQ 7HVW (QYLURQPHQWV G 7UDLQLQJ (QYLURQPHQWV Figure 2: System overview: Given a goal image (a) , RECON explores the environment (b) by sampling prospective latent goals and constructing a topological map of images (white dots), operating only on visual observations. After finding the goal (c) , RECON can reuse the map to reach arbitrary goals in the environment (red path in (b) ). RECON uses data collected from diverse traini g environments (d) to learn navigational priors that enable it to quickly explore and earn to reach vi u l goals a variety of uns en environments (e) . tions. Robustness of this kind is critical in real-world settings, where the appearance of landmarks can change significantly with times of day and seasons of the year. We demonstrate our method, R apid E xploration C ontrollers for O utcome-driven N avigation (RE- CON), on a mobile ground robot (Fig. 1) and evaluate against 6 competitive baselines spanning over 100 hours of real-world experiments in 8 distinct open-world environments (Fig. 2). Our method can discover user-spec fied go ls up to 80m away after just 20 mi utes of inter ction in ew environ- ment. We also demon trat robu tness in th presence of visual distractors and novel bstacles. We make this dataset publicly available as a source of real-world interaction data for future resesarch. 2 Related Work Exploring a new environment is often framed as the problem of efficient mapping, posed in terms of information maximization to guide the robot to uncertain regions of the environment. Some prior exploration methods use local strategies for generating control actions for the robots [1–4], while others use use global strategies based on the frontier method [5–7]. However, building high-fidelity geometric maps can be hard without reliable depth information. Further, such maps do not encode semantic aspects of traversability, e.g., tall grass is traversable but a wire fence is not.