This paper investigates reinforcement learning with constraints, which is indispensable in safety-critical environments. To drive the constraint violation monotonically decrease, the constraints are taken as Lyapunov functions, and new linear constraints are imposed on the updating dynamics of the policy parameters such that the original safety set is forward-invariant in expectation. As the new guaranteed-feasible constraints are imposed on the updating dynamics instead of the original policy parameters, classic optimization algorithms are no longer applicable. To address this, we propose to learn a neural network-based meta-optimizer to optimize the objective while satisfying such linear constraints. The constraint-satisfaction is achieved via projection onto a polytope formulated by multiple linear inequality constraints, which can be solved analytically with our newly designed metric. Ultimately, the meta-optimizer trains the policy network to monotonically decrease the constraint violation and maximize the cumulative reward. Numerical results validate the theoretical findings.
Demand for high-performance, robust, and safe autonomous systems has grown substantially in recent years. Fulfillment of these objectives requires accurate and efficient risk estimation that can be embedded in core decision-making tasks such as motion planning. On one hand, Monte-Carlo (MC) and other sampling-based techniques can provide accurate solutions for a wide variety of motion models but are cumbersome to apply in the context of continuous optimization. On the other hand, "direct" approximations aim to compute (or upper-bound) the failure probability as a smooth function of the decision variables, and thus are widely applicable. However, existing approaches fundamentally assume discrete-time dynamics and can perform unpredictably when applied to continuous-time systems operating in the real world, often manifesting as severe conservatism. State-of-the-art attempts to address this within a conventional discrete-time framework require additional Gaussianity approximations that ultimately produce inconsistency of their own. In this paper we take a fundamentally different approach, deriving a risk approximation framework directly in continuous time and producing a lightweight estimate that actually improves as the discretization is refined. Our approximation is shown to significantly outperform state-of-the-art techniques in replicating the MC estimate while maintaining the functional and computational benefits of a direct method. This enables robust, risk-aware, continuous motion-planning for a broad class of nonlinear, partially-observable systems.
Deep Neural Network-based systems are now the state-of-the-art in many robotics tasks, but their application in safety-critical domains remains dangerous without formal guarantees on network robustness. Small perturbations to sensor inputs (from noise or adversarial examples) are often enough to change network-based decisions, which was recently shown to cause an autonomous vehicle to swerve into another lane. In light of these dangers, numerous algorithms have been developed as defensive mechanisms from these adversarial inputs, some of which provide formal robustness guarantees or certificates. This work leverages research on certified adversarial robustness to develop an online certified defense for deep reinforcement learning algorithms. The proposed defense computes guaranteed lower bounds on state-action values during execution to identify and choose a robust action under a worst-case deviation in input space due to possible adversaries or noise. The approach is demonstrated on a Deep Q-Network policy and is shown to increase robustness to noise and adversaries in pedestrian collision avoidance scenarios and a classic control task. This work extends our previous paper with new performance guarantees, expanded results aggregated across more scenarios, an extension into scenarios with adversarial behavior, comparisons with a more computationally expensive method, and visualizations that provide intuition about the robustness algorithm.
We present a novel POMDP problem formulation for a robot that must autonomously decide where to go to collect new and scientifically relevant images given a limited ability to communicate with its human operator. From this formulation we derive constraints and design principles for the observation model, reward model, and communication strategy of such a robot, exploring techniques to deal with the very high-dimensional observation space and scarcity of relevant training data. We introduce a novel active reward learning strategy based on making queries to help the robot minimize path "regret" online, and evaluate it for suitability in autonomous visual exploration through simulations. We demonstrate that, in some bandwidth-limited environments, this novel regret-based criterion enables the robotic explorer to collect up to 17% more reward per mission than the next-best criterion.
We present Asynchronous Stochastic Parallel Pose Graph Optimization (ASAPP), the first asynchronous algorithm for distributed pose graph optimization (PGO) in multi-robot simultaneous localization and mapping. By enabling robots to optimize their local trajectory estimates without synchronization, ASAPP offers resiliency against communication delays and alleviates the need to wait for stragglers in the network. Furthermore, the same algorithm can be used to solve the so-called rank-restricted semidefinite relaxations of PGO, a crucial class of non-convex Riemannian optimization problems at the center of recent PGO solvers with global optimality guarantees. Under bounded delay, we establish the global first-order convergence of ASAPP using a sufficiently small stepsize. The derived stepsize depends on the worst-case delay and inherent problem sparsity, and furthermore matches known result for synchronous algorithms when delay is zero. Numerical evaluations on both simulated and real-world SLAM datasets demonstrate the speedup achieved with ASAPP and show the algorithm's resilience against a wide range of communication delays in practice.
Disturbance estimation for Micro Aerial Vehicles (MAVs) is crucial for robustness and safety. In this paper, we use novel, bio-inspired airflow sensors to measure the airflow acting on a MAV, and we fuse this information in an Unscented Kalman Filter (UKF) to simultaneously estimate the three-dimensional wind vector, the drag force, and other interaction forces (e.g. due to collisions, interaction with a human) acting on the robot. To this end, we present and compare a fully model-based and a deep learning-based strategy. The model-based approach considers the MAV and airflow sensor dynamics and its interaction with the wind, while the deep learning-based strategy uses a Long Short-Term Memory (LSTM) neural network to obtain an estimate of the relative airflow, which is then fused in the proposed filter. We validate our methods in hardware experiments, showing that we can accurately estimate relative airflow of up to 4 m/s, and we can differentiate drag and interaction force.
Reliance on external localization infrastructure and centralized coordination are main limiting factors for formation flying of vehicles in large numbers and in unprepared environments. While solutions using onboard localization address the dependency on external infrastructure, the associated coordination strategies typically lack collision avoidance and scalability. To address these shortcomings, we present a unified pipeline with onboard localization and a distributed, collision-free motion planning strategy that scales to a large number of vehicles. Since distributed collision avoidance strategies are known to result in gridlock, we also present a decentralized task assignment solution to deconflict vehicles. We experimentally validate our pipeline in simulation and hardware. The results show that our approach for solving the optimization problem associated with motion planning gives solutions within seconds in cases where general purpose solvers fail due to high complexity. In addition, our lightweight assignment strategy leads to successful and quicker formation convergence in 96-100% of all trials, whereas indefinite gridlocks occur without it for 33-50% of trials. By enabling large-scale, deconflicted coordination, this pipeline should help pave the way for anytime, anywhere deployment of aerial swarms.
The complexity of multiagent reinforcement learning (MARL) in multiagent systems increases exponentially with respect to the agent number. This scalability issue prevents MARL from being applied in large-scale multiagent systems. However, one critical feature in MARL that is often neglected is that the interactions between agents are quite sparse. Without exploiting this sparsity structure, existing works aggregate information from all of the agents and thus have a high sample complexity. To address this issue, we propose an adaptive sparse attention mechanism by generalizing a sparsity-inducing activation function. Then a sparse communication graph in MARL is learned by graph neural networks based on this new attention mechanism. Through this sparsity structure, the agents can communicate in an effective as well as efficient way via only selectively attending to agents that matter the most and thus the scale of the MARL problem is reduced with little optimality compromised. Comparative results show that our algorithm can learn an interpretable sparse structure and outperforms previous works by a significant margin on applications involving a large-scale multiagent system.
There are several real-world tasks that would benefit from applying multiagent reinforcement learning (MARL) algorithms, including the coordination among self-driving cars. The real world has challenging conditions for multiagent learning systems, such as its partial observable and nonstationary nature. Moreover, if agents must share a limited resource (e.g. network bandwidth) they must all learn how to coordinate resource use. This paper introduces a deep recurrent multiagent actor-critic framework (R-MADDPG) for handling multiagent coordination under partial observable set-tings and limited communication. We investigate recurrency effects on performance and communication use of a team of agents. We demonstrate that the resulting framework learns time dependencies for sharing missing observations, handling resource limitations, and developing different communication patterns among agents.