Large, pretrained models are commonly finetuned with imagery that is heavily augmented to mimic different conditions and scales, with the resulting models used for various tasks with imagery from a range of spatial scales. Such models overlook scale-specific information in the data for scale-dependent domains, such as remote sensing. In this paper, we present Scale-MAE, a pretraining method that explicitly learns relationships between data at different, known scales throughout the pretraining process. Scale-MAE pre-trains a network by masking an input image at a known input scale, where the area of the Earth covered by the image determines the scale of the ViT positional encoding, not the image resolution. Scale-MAE encodes the masked image with a standard ViT backbone, and then decodes the masked image through a bandpass filter to reconstruct low/high frequency images at lower/higher scales. We find that tasking the network with reconstructing both low/high frequency images leads to robust multiscale representations for remote sensing imagery. Scale-MAE achieves an average of a 2.4 − 5.6% non-parametric kNN classification improvement across eight remote sensing datasets compared to current state-of-the-art and obtains a 0.9 mIoU to 1.7 mIoU improvement on the SpaceNet building segmentation transfer task for a range of evaluation scales.

In this paper, we consider the problem of an autonomous robot searching for a target object whose position is characterized by a prior probability distribution over the workspace (the object prior). We consider the case of a continuous search domain, and a robot equipped with a single binary sensor whose ability to recognize the target object varies probabilistically as a function of the distance from the robot to the target (the sensor model). We show that when the object prior and sensor model are taken from the exponential family of distributions, the searcher's posterior probability map for the object location belongs to a finitely parameterizable class of functions, admitting an exact representation of the searcher's evolving belief. Unfortunately, the cost of the representation grows exponentially with the number of stages in the search. For this reason, we develop an approximation scheme that exploits regularized particle filtering methods. We present simulation studies for several scenarios to demonstrate the effectiveness of our approach using a simple, greedy search strategy.

In this paper, we present a hierarchical framework that enables motion planning for asymptotically stable 3-D bipedal walking in the same way that planning is already possible for zero moment point walking. This framework is based on the construction of asymptotically stable gait primitives for a class of hybrid dynamical systems with impacts. Each primitive corresponds to an asymptotically stable hybrid limit cycle that admits rules a priori for sequential composition with other primitives, reducing a high-dimensional feedback motion planning problem into a low-dimensional discrete tree search. As a constructive example, we develop this planning framework for the 3-D compass-gait biped, where each primitive corresponds to walking along an arc of constant curvature for a fixed number of steps. We apply a discrete search algorithm to plan a sequence of these primitives, taking the 3-D biped stably from start to goal in workspaces with obstacles. We finally show how this framework generalizes to more complex models by planning walking paths for an underactuated five-link biped.

A ubiquitous problem in robotics is determining policies that move robots with uncertain process and observation models (partially-observed state systems) to a goal configuration while avoiding collision. We propose a new method to solve this minimum uncertainty navigation problem. We use a continuous partially-observable Markov decision process (POMDP) model and optimize an objective function that considers both probability of collision and uncertainty at the goal position. By using information-theoretic heuristics, we are able to find policies that are effective for both minimizing collisions and stopping near the goal configuration. We additionally introduce a filtering algorithm that tracks collision free trajectories and estimates the probability of collision.

We propose a planning algorithm that allows user-supplied domain knowledge to be exploited in the synthesis of information feedback policies for systems modeled as partially observable Markov decision processes (POMDPs). POMDP models, which are increasingly popular in the robotics literature, permit a planner to consider future uncertainty in both the application of actions and sensing of observations. With our approach, domain experts can inject specialized knowledge into the planning process by providing a set of local policies that are used as primitives by the planner. If the local policies are chosen appropriately, the planner can evaluate further into the future, even for large problems, which can lead to better overall policies at decreased computational cost. We use a structured approach to encode the provided domain knowledge into the value function approximation. We demonstrate our approach on a multi-robot fire fighting problem, in which a team of robots cooperates to extinguish a spreading fire, modeled as a stochastic process. The state space for this problem is significantly larger than is typical in the POMDP literature, and the geometry of the problem allows for the application of an intuitive set of local policies, thus demonstrating the effectiveness of our approach.

We propose a new minimum uncertainty planning technique for mobile robots localizing with beacons. We model the system as a partially-observable Markov decision process and use a sampling-based method in the belief space (the space of posterior probability density functions over the state space) to find a belief-feedback policy. This approach allows us to analyze the evolution of the belief more accurately, which can result in improved policies when common approximations do not model the true behavior of the system. We demonstrate that our method performs comparatively, and in certain cases better, than current methods in the literature.