We present a new approach to learning sparse, spatiotemporal codes in which the number of basis vectors, their orientations, velocities and the size of their receptive fields change over the duration of unsupervised training. The algorithm starts with a relatively small, initial basis with minimal temporal extent. This initial basis is obtained through conventional sparse coding techniques and is expanded over time by recursively constructing a new basis consisting of basis vectors with larger temporal extent that proportionally conserve regions of previously trained weights. These proportionally conserved weights are combined with the result of adjusting newly added weights to represent a greater range of primitive motion features. The size of the current basis is determined probabilistically by sampling from existing basis vectors according to their activation on the training set. The resulting algorithm produces bases consisting of filters that are bandpass, spatially oriented and temporally diverse in terms of their transformations and velocities. The basic methodology borrows inspiration from the layer-by-layer learning of multiple-layer restricted Boltzmann machines developed by Geoff Hinton and his students. Indeed, we can learn multiple-layer sparse codes by training a stack of denoising autoencoders, but we have had greater success using L1regularized regression in a variation on Olshausen and Field's original SPARSENET. To accelerate learning and focus attention, we apply a space-time interest-point operator that selects for periodic motion. This attentional mechanism enables us to efficiently compute and compactly represent a broad range of interesting motion. We demonstrate the utility of our approach by using it to recognize human activity in video. Our algorithm meets or exceeds the performance of state-of-the-art activity-recognition methods.

In this paper, we consider the problem of evaluating plan fragments for insertion into an existing plan. Preparing a plan that is robust to variations in continuous resources requires considering the uncertainty over these resources. For some situations, it is impractical to add branches to the plan a priori, because such situations may arise at any of a large number of points in execution. Adding all these potential branches to an existing plan would require an immense computational effort and result in a huge plan. In such a case, it makes more sense to precompute a plan fragment to react to such a situation and allow the execution system to insert the plan fragment into its current plan when appropriate. The problem can thus be seen as a restricted form of plan merging, but with the addition of uncertainty and continuous resources, and the criterion of maximizing utility rather than absolute goal achievement. We formalize the problem and introduce an efficient solution method, incremental bounds refinement, that iteratively narrows bounds on the expected value of adding a branch until it can be discriminated from the original plan.

Future Mars rovers, such as the planned 2009 MSL rover, require sufficient autonomy to robustly approach rock targets and place an instrument in contact with them. It took the 1997 Sojourner Mars rover between 3 and 5 communications cycles to accomplish this. This paper describes the NASA Ames approach to robustly accomplishing single cycle instrument deployment, using the K9 prototype Mars rover. An off-board 3D site model is used to select science targets for the rover. K9 navigates to targets using deduced reckoning, and autonomously assesses the target area to determine where to place an arm mounted microscopic camera. Onboard K9 is a resource cognizant conditional executive, which extends the complexity and duration of operations that a can be accomplished without intervention from mission control.

A planetary rover must be able to identify states where it should stop or change its plan. With limited and infrequent communication from ground, the rover must recognize states accurately. However, the sensor data is inherently noisy, so identifying the temporal patterns of data that correspond to interesting or important states becomes a complex problem. We present an approach to state identification using second-order hidden Markov models. Models are trained automatically on a set of labeled training data; the rover uses those models to identify its state from the observed data. The approach is demonstrated on data from a planetary rover platform.

For extended autonomous operation, rovers must identify potential faults to determine whether its execution needs to be halted or not. At the same time, rovers present particular challenges for state estimation techniques: they are subject to environmental influences that affect sensor readings during normal and anomalous operation, and the sensors fluctuate rapidly due to both the noise and the dynamics of rover's interaction with its environment. This paper presents MAKSI, an on-board method for state estimation and fault diagnosis that is particularly appropriate for rovers. The method is based on a combination of continuous state estimation, using Kalman filters, and discrete state estimation using Markov-model representation.

This paper presents an approach to building plans using partially observable Markov decision processes. The approach begins with a base solution that assumes full observability. The partially observable solution is incrementally constructed by considering increasing amounts of information from observations. The base solution directs the expansion of the plan by providing an evaluation function for the search fringe. We show that incremental observation moves from the base solution towards the complete solution, allowing the planner to model the uncertainty about action outcomes and observations that are present in real domains.

Effective monitoring of device-supported patients in the intensive care unit (ICU) is complex, involving interpretation of many variables, comparative evaluation of many therapy options, and control of many patient-management parameters. Even skilled clinicians make errors that limit the quality of care, harm patients, or cause life-threatening situations. A growing body of research aims to improve ICU monitoring with computer technology. Most of this research falls in two areas: (a) short-term engineering of practical solutions to narrowly defined immediate problems (e.g. smart alarm systems); or (b) basic research on fundamental issues potentially relevant to ICU monitoring (e.g. temporal reasoning). By contrast, our project aims to develop a more comprehensive ‘intelligent agent’, having a broad range of capabilities, to cooperate on the ICU team. We do not aim to produce a practical system suitable for near-term deployment in the ICU, but rather a ‘proof of concept’, an experimental system that: (a) demonstrably performs and coordinates a range of intelligent reasoning tasks of use in ICU monitoring; (b) does so reliably in a significant range of medical situations; and (c) arguably will scale up to meet the comprehensive set of practical requirements with an appropriate development effort. We have developed an experimental system called Guardian, which exhibits several of the required capabilities and utilizes an underlying architecture hypothesized to support the full range of required capabilities. In this paper, we describe the Guardian system, its architecture, and its current knowledge base. We describe its performance and summarize the results of preliminary evaluations. Finally, we discuss ongoing and planned research on Guardian.