Souma Chowdhury

Off-road navigation is critical for a wide range of field robotics applications from planetary exploration to disaster response. However, it remains a longstanding challenge due to unstructured environments and the inherently complex terrain–vehicle interactions. Traditional physics-based methods struggle to accurately capture the nonlinear dynamics underlying these interactions, while purely data-driven approaches often overfit to specific motion patterns, vehicle geometries, or platforms, limiting their generalization in diverse, real-world scenarios. To address these limitations, we introduce AnyNav, a vision-based friction estimation and navigation framework grounded in neuro-symbolic principles. Our approach integrates neural networks for visual perception with symbolic physical models for reasoning about terrain-vehicle dynamics. To enable learning in a self-supervised manner for real-world settings, we adopt the imperative learning paradigm, where a bilevel optimization framework leverages physics-based optimization for training the friction network. This explicit incorporation of physical reasoning substantially improves generalization across terrains, vehicle types, and operating conditions. Leveraging the predicted friction coefficients, we further develop a physics-informed navigation system capable of generating physically feasible, time-efficient paths together with corresponding speed profiles. We demonstrate that AnyNav seamlessly transfers from simulation to real-world robotic platforms, exhibiting strong robustness across different four-wheeled vehicles and diverse off-road environments.

Active tether-net systems offer a promising solution to the capture and disposal of large space debris that pose significant safety risks to the operation of satellites in Earth's orbit. These systems comprise a flexible net with maneuverable units (MUs) launched from a small chaser spacecraft, with the MUs assumed to have independent thrust capabilities, such as based on cold gas propulsion. The requirement of autonomous operation of this system that must also work (generalize) across varying relative state of the target debris makes the capture process challenging. Given these requirements and the complexity of modeling the deployment dynamics of the net and net/debris interactions, this paper presents neural-network-based policies to guide the maneuvering (thrust) actions of the MUs, which are trained using reinforcement learning (RL). To this end, the tether-net maneuver operation is modeled as a Markov Decision Process with the state encompassing the target's location, orientation and rotation rates relative to the chaser, and solved using a policy gradient approach. A special reward function is formulated to account for both the primary objectives of capture success metrics and fuel consumptions incurred by the MUs, and intermediate objectives related to hypothesized favorable intermediate states of the net that facilitate capture performance and alleviate reward sparsity during RL training. To avoid the otherwise prohibitive computing cost of training the maneuver policy via RL, surrogate models, respectively based on Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), are trained for predicting the deployment trajectory and capture outcome, and used in various combinations with a higher-fidelity simulation as the RL environment. Significant improvement in capture performance, from around 5.7% (with random actions) to 85% is achieved via training. Tested over unseen scenarios, the best RL-derived policy readily outperforms an aiming-point-based baseline, with the nature of actions providing important insights into net behavior (e.g., active net rotation and maximal mouth opening) favorable to capture success and low fuel consumption by MUs.

Advanced manufacturing processes are often based on complex multiphysics phenomena that are either poorly understood or are computationally too expensive to simulate in the context of process design, control, or planning. Traditionally, simplified physics models with prescribed heuristics or purely data-driven surrogate models are used as alternatives in such applications. The concept of physics-informed machine learning (PIML) has been shown to have unique advantages over both of these alternatives in various fields of complex system analysis. In this paper, a new PIML approach is presented to model the geometry of the cut produced by a magnetically assisted laser-induced plasma micro-machining (M-LIPMM) process. This PIML architecture uses a neural network to auto-adapt the parametric boundary condition and physical properties used in a simplified finite difference-based physics model (of 2D heat conduction), as a function of the inputs namely the laser settings. This network also estimates the scaling and shifting parameters used by a convolutional neural network that takes the temperature profile predicted by the simplified heat conduction model to predict the width and depth of the machined cut. Trained on physical experiment data, the PIML approach compares favorably to a pure data-driven neural network model in extrapolation tests, while also providing physical insights (that the latter cannot). The PIML approach also provides an 85% better accuracy overall compared to the simplified physics model with heuristic settings.

Interesting and efficient collective behavior observed in multi-robot or swarm systems emerges from the individual behavior of the robots. The functional space of individual robot behaviors is in turn shaped or constrained by the robot's morphology or physical design. Thus the full potential of multi-robot systems can be realized by concurrently optimizing the morphology and behavior of individual robots, informed by the environment's feedback about their collective performance, as opposed to treating morphology and behavior choices disparately or in sequence (the classical approach). This paper presents an efficient concurrent design or co-design method to explore this potential and understand how morphology choices impact collective behavior, particularly in an MRTA problem focused on a flood response scenario, where the individual behavior is designed via graph reinforcement learning. Computational efficiency in this case is attributed to a new way of near exact decomposition of the co-design problem into a series of simpler optimization and learning problems. This is achieved through i) the identification and use of the Pareto front of Talent metrics that represent morphology-dependent robot capabilities, and ii) learning the selection of Talent best trade-offs and individual robot policy that jointly maximizes the MRTA performance. Applied to a multi-unmanned aerial vehicle flood response use case, the co-design outcomes are shown to readily outperform sequential design baselines. Significant differences in morphology and learned behavior are also observed when comparing co-designed single robot vs. co-designed multi-robot systems for similar operations.

Legged locomotion shows promise for running in complex, unstructured environments. Designing such legged robots requires considering heterogeneous, multi-domain constraints and variables, from mechanical hardware and geometry choices to controller profiles. However, very few formal or systematic (as opposed to ad hoc) design formulations and frameworks exist to identify feasible and robust running platforms, especially at the small (sub 500 g) scale. This critical gap in running legged robot design is addressed here by abstracting the motion of legged robots through a torque-driven spring-loaded inverted pendulum (TD-SLIP) model, and deriving constraints that result in stable cyclic forward locomotion in the presence of system noise. Synthetic noise is added to the initial state in candidate design evaluation to simulate accumulated errors in an open-loop control. The design space was defined in terms of morphological parameters, such as the leg properties and system mass, actuator selection, and an open loop voltage profile. These attributes were optimized with a well-known particle swarm optimization solver that can handle mixed-discrete variables. Two separate case studies minimized the difference in touchdown angle from stride to stride and the actuation energy, respectively. Both cases resulted in legged robot designs with relatively repeatable and stable dynamics, while presenting distinct geometry and controller profile choices.

Self-healing smart grids are characterized by fast-acting, intelligent control mechanisms that minimize power disruptions during outages. The corrective actions adopted during outages in power distribution networks include reconfiguration through switching control and emergency load shedding. The conventional decision-making models for outage mitigation are, however, not suitable for smart grids due to their slow response and computational inefficiency. Here, we present a graph reinforcement learning model for outage management in the distribution network to enhance its resilience. The distinctive characteristic of our approach is that it explicitly accounts for the underlying network topology and its variations with switching control, while also capturing the complex interdependencies between state variables (along nodes and edges) by modeling the task as a graph learning problem. Our model learns the optimal control policy for power restoration using a Capsule-based graph neural network. We validate our model on three test networks, namely the 13, 34, and 123-bus modified IEEE networks where it is shown to achieve near-optimal, real-time performance. The resilience improvement of our model in terms of loss of energy is 607.45 kWs and 596.52 kWs for 13 and 34 buses, respectively. Our model also demonstrates generalizability across a broad range of outage scenarios.

Optimal decision-making is key to efficient allocation and scheduling of repair resources (e.g., crews) to service affected nodes of large power grid networks. Traditional manual restoration methods are inadequate for modern smart grids sprawling across vast territories, compounded by the unpredictable nature of damage and disruptions in power and transportation networks. This paper develops a method that focuses on the restoration and repair efforts within power systems. We expand upon the methodology proposed in the literature and incorporate a real-world transportation network to enhance the realism and practicality of repair schedules. Our approach carefully devises a reduced network that combines vulnerable components from the distribution network with the real transportation network. Key contributions include dynamically addressing a coupled resource allocation and capacitated vehicle routing problem over a new reduced network model, constructed by integrating the power grid with the transportation network. This is performed using network heuristics and graph theory to prioritize securing critical grid segments. A case study is presented for the 8500 bus system.