Working memory (WM) is a cognitive function for temporary maintenance and manipulation of information, which requires conversion of stimulus-driven signals into internal representations that are maintained across seconds-long mnemonic delays. Within primate prefrontal cortex (PFC), a critical node of the brain's WM network, neurons show stimulus-selective persistent activity during WM, but many of them exhibit strong temporal dynamics and heterogeneity, raising the questions of whether, and how, neuronal populations in PFC maintain stable mnemonic representations of stimuli during WM. Here we show that despite complex and heterogeneous temporal dynamics in single-neuron activity, PFC activity is endowed with a population-level coding of the mnemonic stimulus that is stable and robust throughout WM maintenance. We applied population-level analyses to hundreds of recorded single neurons from lateral PFC of monkeys performing two seminal tasks that demand parametric WM: oculomotor delayed response and vibrotactile delayed discrimination. We found that the high-dimensional state space of PFC population activity contains a low-dimensional subspace in which stimulus representations are stable across time during the cue and delay epochs, enabling robust and generalizable decoding compared with time-optimized subspaces. To explore potential mechanisms, we applied these same population-level analyses to theoretical neural circuit models of WM activity. Three previously proposed models failed to capture the key population-level features observed empirically. We propose network connectivity properties, implemented in a linear network model, which can underlie these features. This work uncovers stable population-level WM representations in PFC, despite strong temporal neural dynamics, thereby providing insights into neural circuit mechanisms supporting WM.working memory | prefrontal cortex | population coding T he neuronal basis of working memory (WM) in prefrontal cortex (PFC) has been studied for decades through singleneuron recordings from monkeys performing tasks in which a transient sensory stimulus must be held in WM across a secondslong delay to guide a future response. These studies discovered that a key neural correlate of WM in PFC is stimulus-selective persistent activity, i.e., stable elevated firing rates in a subset of neurons, that spans the delay (1). These neurophysiological findings have grounded a leading hypothesis that WM is supported by stable persistent activity patterns in PFC that bridge the gap between stimulus and response epochs. Because the timescales of WM maintenance (several seconds) are longer than typical timescales of neuronal and synaptic integration (∼10-100 ms), mechanisms at the level of neural circuits may be critical for generating WM activity in PFC (2). A leading theoretical framework proposes that PFC circuits subserve WM maintenance through dynamical attractors, i.e., stable fixed points in network activity, generated by strong recurrent connectivity (3, 4).Recent neurophysiologi...

Abstract-In this paper we address the problem of motion planning in the presence of state uncertainty, also known as planning in belief space. The work is motivated by planning domains involving nontrivial dynamics, spatially varying measurement properties, and obstacle constraints. To make the problem tractable, we restrict the motion plan to a nominal trajectory stabilized with a linear estimator and controller. This allows us to predict distributions over future states given a candidate nominal trajectory. Using these distributions to ensure a bounded probability of collision, the algorithm incrementally constructs a graph of trajectories through state space, while efficiently searching over candidate paths through the graph at each iteration. This process results in a search tree in belief space that provably converges to the optimal path. We analyze the algorithm theoretically and also provide simulation results demonstrating its utility for balancing information gathering to reduce uncertainty and finding low cost paths.

Abstract-When a mobile agent does not known its position perfectly, incorporating the predicted uncertainty of future position estimates into the planning process can lead to substantially better motion performance. However, planning in the space of probabilistic position estimates, or belief space, can incur substantial computational cost. In this paper, we show that planning in belief space can be done efficiently for linear Gaussian systems by using a factored form of the covariance matrix. This factored form allows several prediction and measurement steps to be combined into a single linear transfer function, leading to very efficient posterior belief prediction during planning. We give a belief-space variant of the Probabilistic Roadmap algorithm called the Belief Roadmap (BRM) and show that the BRM can compute plans substantially faster than conventional belief space planning. We conclude with performance results for an agent using ultra-wide bandwidth (UWB) radio beacons to localize and show that we can efficiently generate plans that avoid failures due to loss of accurate position estimation.

Classical models of perceptual decision-making assume that animals use a single, consistent strategy to form decisions, or that decision-making strategies evolve slowly over time. Here we present new analyses suggesting that this common view is incorrect. We analyzed data from two mouse decision-making experiments and found that choice behavior relies on an interplay between multiple interleaved strategies. These strategies, characterized by states in a hidden Markov model, persist for tens to hundreds of trials before switching, and may alternate multiple times within a session.The identified strategies were highly consistent across animals, consisting of a single "engaged" state, in which decisions relied heavily on the sensory stimulus, and several biased or disengaged states in which errors frequently occurred. These results provide a powerful alternate explanation for "lapses" often observed in psychophysical experiments, and suggest that standard measures of performance mask the presence of dramatic changes in strategy across trials.

This video highlights our system that enables a Micro Aerial Vehicle (MAV) to autonomously explore and map unstructured and unknown GPS-denied environments. While mapping and exploration solutions are now well-established for ground vehicles, air vehicles face unique challenges which have hindered the development of similar capabilities. Although there has been recent progress toward sensing, control, and navigation techniques for GPS-denied flight, there have been few demonstrations of stable, goal-directed flight in real-world environments. Our system leverages a multi-level sensing and control hierarchy that matches the computational complexity of the component algorithms with the real-time needs of a MAV to achieve autonomy in unconstrained environments.

Abstract-Robots that use learned perceptual models in the real world must be able to safely handle cases where they are forced to make decisions in scenarios that are unlike any of their training examples. However, state-of-the-art deep learning methods are known to produce erratic or unsafe predictions when faced with novel inputs. Furthermore, recent ensemble, bootstrap and dropout methods for quantifying neural network uncertainty may not efficiently provide accurate uncertainty estimates when queried with inputs that are very different from their training data. Rather than unconditionally trusting the predictions of a neural network for unpredictable real-world data, we use an autoencoder to recognize when a query is novel, and revert to a safe prior behavior. With this capability, we can deploy an autonomous deep learning system in arbitrary environments, without concern for whether it has received the appropriate training. We demonstrate our method with a vision-guided robot that can leverage its deep neural network to navigate 50% faster than a safe baseline policy in familiar types of environments, while reverting to the prior behavior in novel environments so that it can safely collect additional training data and continually improve. A video illustrating our approach is available at: http://groups.csail.mit.edu/rrg/videos/safe visual navigation.

No abstract

This paper presents our solution for enabling a quadrotor helicopter, equipped with a laser rangefinder sensor, to autonomously explore and map unstructured and unknown indoor environments. While these capabilities are already commodities on ground vehicles, air vehicles seeking the same performance face unique challenges. In this paper, we describe the difficulties in achieving fully autonomous helicopter flight, highlighting the differences between ground and helicopter robots that make it difficult to use algorithms that have been developed for ground robots. We then provide an overview of our solution to the key problems, including a multilevel sensing and control hierarchy, a high-speed laser scan-matching algorithm, an EKF for data fusion, a high-level SLAM implementation, and an exploration planner. 1 Finally, we show experimental results demonstrating the helicopter's ability to navigate accurately and autonomously in unknown environments. INTRODUCTIONMicro Aerial Vehicles (MAVs) are increasingly being used in military and civilian domains, including surveillance operations, weather observation, and disaster relief coordination. Enabled by GPS and MEMS inertial sensors, MAVs that can fly in outdoor environments without human intervention have been developed [2,3,4,5].Unfortunately, most indoor environments and many parts of the urban canyon remain without access to external positioning systems such as GPS. Autonomous MAVs today are thus limited in their ability to fly through these areas. Traditionally, unmanned vehicles operating in GPS-denied environments can rely on dead reckoning for localization, but these measurements drift over time. Alternatively, simultaneous localization and mapping (SLAM) algorithms build a map of the environment around the vehicle while simultaneously using it to estimate the vehicle's position. Although there have been significant advances in developing accurate, drift-free SLAM algorithms for large-scale environments, these algorithms have focused almost exclusively on ground or underwater vehicles. In contrast, attempts to achieve the same results with MAVs have not been as successful due to a combination of limited payloads for sensing and computation, coupled with the fast, unstable dynamics of the air vehicles.In this work, we present our quadrotor helicopter system, shown in Figure 1, that is capable of autonomous flight in unstructured indoor environments, such as the one shown in Figure 2. The system employs a multi-level sensor processing hierarchy designed to meet the requirements for controlling a helicopter. The key contribution of this paper is the development of a fully autonomous quadrotor that relies only on onboard sensors for stable control without requiring prior maps of the environment.After discussing related work in Section 2, we begin in Section 3 by analyzing the key challenges MAVs face when attempting to perform SLAM. We then give an overview of the algorithms employed by our system. Finally, we demonstrate our helicopter navigating autonomo...