Arne Suppé's research while affiliated with Carnegie Mellon University and other places

... Using deep learning, recognition can bridge the gap between perception and intelligence. For example, perception missions can collect lots of data, but much of it was discarded to simplify interpretation by higher level tasks [20], e.g. a 3D object can become a point in space. In Arne's paper [20], it used a deep learning framework to replace these interfaces with learned interfaces. ...

... Biologically inspired multi-legged robots hold the promise for robustly traversing cluttered terrain, with inherent advantage of legs in overcoming large obstacles (Raibert, 2008) and higher stability offered by multiple legs (Full et al., 2006;Ting et al., 1994). By exploiting natural dynamics of leg-ground interaction (Childers et al., 2016;De and Koditschek, 2018;Garcia Bermudez et al., 2012;Li et al., 2009;Qian and Koditschek, 2020;Raibert, 2008;Saranli et al., 2001;Spagna et al., 2007) and body-terrain interaction Li et al., 2015), multi-legged robots have achieved improved locomotor performance and capabilities in complex terrain compared to those using quasi-static planning and control (although still far from robust (Arslan and Saranli, 2012)). For example, a recent study empirically discovered that an ellipsoidal body shape helps traverse cluttered obstacles by inducing body rolling through gaps narrower than the body width (Li et al., 2015). ...

... The arrival rate measures how often an agent reaches its a goal that is a primary metric in general robot navigation [167] and has also been used in social navigation [231]. In general, some assumptions are added to measure the arrival rate. ...

... In our prior work, we have demonstrated approaches to reduce the time and effort required to assign labels to visual data and adapt visual classifiers to novel environments (Wigness et al. 2016; Wigness and Rogers III 2017). This paper describes the combination of these techniques with a method for learning robot control parameters through inverse optimal control (IOC). ...

... They typically have to figure out where they are, navigate through foreign areas, and ask people for directions to their destination. State-of-the-art artificial-intelligence and humanrobot-interaction research enables robots to solve such naturallanguage-driven navigation tasks [3,5,15,16,19,21,22,22,30]. However, these systems suffer from certain limitations such as requiring a specific syntactic structure for naturallanguage commands or requiring a map of the environment. ...

... An earlier work on the idea of using human teammates in robot perception was briefly introduced in [16]. The temporal update method used, however, was too domain-specific to represent general dependence relationships among variables, thus lacking flexibility of generalization. ...

... The importance of curb/step detection for autonomous mobility is well established in literature, with a diverse set of sensors (such as in [12]) being used for the purpose. In [13], an active range-finder employing a laser and camera system is described that uses Jump-Markov modeling to track distances to the ground plane and detect steps. ...

... They only look at targets inside the regions of interest and use the detections to keep a safe distance to the obstacles. Finally, Mertz et al. [172] proposed a detection and tracking framework relying on multiple lidar sources that can be 2D or 3D. Using a combination of segments and features in the point cloud, they associated the measurements to obstacles and keep updated a list of current obstacles. ...

... Scanning LADARs have traditionally been used for mapping and analyzing stationary objects, (see [2], [3], [4], [5], [6], and [7]). Recently, there have been a number of extensions to detect moving targets using techniques such as global segmentation, [8], cluster similarity, [9], feature detection, [10], [11], [12], [13], [14], [15], [16], and model fitting, [17]. Of the techniques that explicitly seek to detect vehicles, most rely on finding straight-edge features in the data and infer vehicle positions from one or more of these edges (see [11], [12], [13], [14] and [15]). ...

... To date, research on parking systems, a type of advanced driver assistance system (ADAS), spans varying degrees of parking assistance including attentional cueing [1], multimodal displays [2]- [5], parking spot identification [6], and fully autonomous parking [7]- [9]. Automated parking is a necessary partially automated safety feature for moving towards a high level of vehicle automation in the future (i.e., Level 4 or Level 5 automation). ...