Benjamin Longmier

Pleased to report that our latest work GRAM: Generalization in Deep RL With a Robust Adaptation Module is now available via RA-L. Congrats to Jimmy Queeney for getting this published. https://lnkd.in/e3z-WsMY Abstract: The reliable deployment of deep reinforcement learning in real-world settings requires the ability to generalize across a variety of conditions, including both in-distribution scenarios seen during training as well as novel out-of-distribution scenarios. In this work, we present a framework for dynamics generalization in deep reinforcement learning that unifies these two distinct types of generalization within a single architecture. We introduce a robust adaptation module that provides a mechanism for identifying and reacting to both in-distribution and out-of-distribution environment dynamics, along with a joint training pipeline that combines the goals of in-distribution adaptation and out-of-distribution robustness. Our algorithm GRAM achieves strong generalization performance across in-distribution and out-of-distribution scenarios upon deployment, which we demonstrate through extensive simulation and hardware locomotion experiments on a quadruped robot.

72

I’ve been reading Arthur C. Clarke’s The Exploration of Space (published in 1951) and it’s incredible how relevant it still feels to modern spacecraft design. Clarke intuitively broke vehicles into the same subsystems we integrate today—propulsion, GNC, structures, thermal, life support—and understood how tightly coupled they must be to succeed. The chapters that explain the need from staging and complex fuel transfer are beautifully explained.  For anyone working in space engineering, it’s a humbling reminder that the challenges we tackle now were already visible to those who looked at the problem with first principles in hand and imagination. Now let's go tackle them!

275

6 Comments

As autonomous driving expands into public road networks, recent examples like the one described in the post below, again demonstrate a critical insight: vehicle intelligence alone is not enough — we need intelligent, responsive road side infrastructure to complement it. At DiMOS Operations GmbH, we are building exactly that. Our focus lies on infrastructure-supported automated driving, with a strong emphasis on external dependencies like Positioning, Navigation & Timing (PNT) and connectivity, which are vulnerable to jamming, spoofing, and outages. To address this, we are launching the first version of our Road Side Early Warning Service (EWS) within 2026, providing real-time PNT integrity monitoring and threat detection for highway vehicle operations. This is a key enabler for scalable, safe, and certifiable deployment of autonomous functions — both in commercial vehicles and passenger cars. #InfrastructureSupportedDriving #LEOPNT #PNT #Integrity #AutonomousDriving #ConnectedVehicles #SafetyOfLife

14

1 Comment

CDI Awarded DOE SBIR Phase I Effort for the “Development of a Modular Multiphysics Simulation Platform for Aerosol and Particulate Dispersion Modeling” CDI has been awarded a Phase I SBIR grant from the US Department of Energy to develop a multiphysics modeling framework for predicting the long-term formation, dynamics, and dispersion of aerosols and particulates released behind aircraft that comprise the nuclei for water and ice accumulation. The effort builds upon legacy work at CDI (see image) to produce a peta- and exa-scale analysis tool that can simulate the chemistry, phase mechanics, vortex dynamics, and turbulence occurring in the wake of a complex-geometry aircraft and assess their impact on the generation of soot and other pollutants in the atmosphere and the formation and persistence of contrails. This capability can, in turn, be used to improve aircraft operations as well as aircraft and aircraft engine design to mitigate the adverse effects associated with released particulates and contrails. To accurately model and resolve these complex dynamical interactions, an adaptive Cartesian grid-based Navier-Stokes flow solver CGE will be used in the vicinity of the flight vehicle. CGE employs a cut-cell mesh generation algorithm to automatically assemble an adaptive volume mesh about the complex and multi-component surfaces (e.g., wingtips, winglets, flaps, etc.) of modern aircraft and propulsion units (turboprops, turbofans). The downstream region containing the wake and particulate dynamics is represented using an advanced block structured code based on DOE’s Flash-X and AMReX frameworks to conduct multiphysics simulations on advanced high performance computational hardware.  

49

2 Comments

Despite maintaining a relatively low profile since joining UCF, my research group has made progress across multiple areas. Attached is a snapshot of real-time data acquisition from our latest test rig, designed for VTOL rotor and propeller experiments. While I’ll leave its specific purposes and applications open for speculation, I welcome discussions from those intrigued by potential collaborations. We look forward to sharing exciting results next year in West Palm Beach—stay tuned!

31

For those of you at #AIAASciTech don't miss our paper on an open source photorealistic simulation and associated tools that help explore solutions to the collision avoidance problem, an area of active research in #Autonomy. The presentation is in Bayhill 28 at 3:30 p.m. by Michael Acheson You can find our paper at https://lnkd.in/eJ8ZqFZB Abstract: Autonomous collision avoidance (ACA) for aerospace systems is an area of active study within the autonomy research community. This problem consists of three consecutive subtasks: perception, prediction, and planning. Each of these sub-tasks are large, active fields of research on their own. For this reason, this team developed an open-source simulation frame work that enables the autonomy community to target ACA using a multirotor vehicle which must avoid moving baseballs. The Baseball Avoidance Multirotor (BAM) simulation framework uses MATLAB® and Simulink® with a multirotor dynamics model in combination with the high-fidelity photo-realistic and sensor simulation capabilities of Microsoft∗ AirSim or Colosseum with Unreal® Engine. The multirotor was selected as the aerospace vehicle due to its simple dynamics and control methodologies that are accessible, high precision, and well known within academia and industry. These features allow researchers to focus on one or more ACA subtasks without managing the more challenging aspects of vehicle control. This paper presents the BAM simulation and details a series of collision avoidance challenges provided with accompanying datasets. Results from BAM simulation experiments demonstrate collision avoidance with both an algorithm that combines Optimal Reciprocal Collision Avoidance (ORCA) with Bernstein polynomials, and a transformer-based replanning method

18

Quantum gravity U(1)-gauge theory renormalizes UAP mass using virtual dark photons that only interact with fermion quantum mass number and not with fermion electric charge, vastly reducing the energy requirement to mimic UAP flight dynamics. Therefore, we hope that Dr. Eric Davis is wrong about the energy requirement to mimic UAP flight dynamics. We use the well-known quantum gravity U(1)-gauge theory from India to mimic UAP flight dynamics on paper, that has a much smaller energy requirement. https://lnkd.in/gpJHMpvt  

6

The term Modus Operandus is a Latin term that can be translated as: 1) Method of Operating. or 2) Way of Behaving. I maintain the Modus Operandus is an important pretext for any system. Especially those systems we humans wish to design, verify, & field into operation. The need to clearly define (via modeling) the Modus Operandi for any system we wish to develop is the foundational basis for what I call the System Operational Architecture. This is one system dimension within my 3D Orthogonal (3DO) architecture framework. An Operational Architecture Model, as 3DO defines it, is a very rich semantic information model for any system. It is based primarily on Mission, Operational Activity and Operational Flow Items as distinct, major ontology classes. But other classes like Metrics, Risks, Interfaces, Standards, and Requirements all cut across any system's operational architecture.

13

5 Comments

Over the past week, I’ve teamed up with Shane Stamm to offer mechanical systems consulting and contract engineering services focused on deployable mechanisms and autonomous docking systems for space applications. Shane brings an exceptional background in on‑orbit servicing and robotic capture mechanisms. As Chief Engineer for the Starsys docking system developed for DARPA’s Orbital Express mission, he led the engineering of one of the earliest and most successful demonstrations of autonomous spacecraft capture. For anyone interested in the technical heritage behind this work, I highly recommend the SPIE paper he co‑authored: “Orbital Express Capture System: Concept to Reality” https://lnkd.in/gSJr4XBT Together, Shane and I provide deep experience across: - Autonomous docking & capture systems - Deployable mechanical systems - Robotic interfaces and alignment mechanisms - Structural, kinematic, and mechanism development for spaceflight - Early‑phase concept development through flight‑ready hardware If your team is developing servicing spacecraft, in‑space robotic systems, or deployable mechanical architectures, we’d be happy to connect and discuss how we can support your mission. Feel free to reach out directly!

16

Excited to highlight "Predicting airfoil dynamic stall loads using neural networks," published in Aerospace Science and Technology (https://lnkd.in/eTCAXCxC) that shows a simple feedforward neural network (FNN) accurately predicting the full aerodynamic cycle - drag, lift, and pitching moment - for a NACA0012 airfoil in dynamic stall (Re ≈ 1.4×10^6). The FNN, trained on Green & Giuni's experimental database of 223 sinusoidal pitching cases, uses the angle of attack and pitch rate as inputs for phase-locked outputs with aggregate R² ≥ 0.99 and ≥ 0.90 per case across light-to-deep stall. Key advantages include computational simplicity without recurrent layers, instant whole-cycle predictions, robust handling of experimental noise, and strong generalization via efficient Octave implementation (open-source). It captures hysteresis, peaks, and trends even in challenging high angles of attack with vortex shedding, smoothing local fluctuations as a deterministic model. AI neural networks like this FNN will completely overtake traditional modeling, slashing CFD costs for many application, for instance, flutter/stability in rotors, UAVs, and wind turbines - enabling real-time, nonlinear models that legacy methods cannot match. #Aerospace #AI #NeuralNetworks #DynamicStall #Engineering #OpenSource

14

2 Comments