Reward-based model tuning becomes provable Newton steps
Fine-tuning and sampling share one iteration with global descent and quadratic convergence.
· “Newton Matching for Generative Modeling: A Unified Framework for Fine-Tuning and Sampling”
eess.SY
Systems and Control
This section includes theoretical and experimental research covering all facets of automatic control systems. The section is focused on methods of control system analysis and design using tools of modeling, simulation and optimization. Specific areas of research include nonlinear, distributed, adaptive, stochastic and robust control in addition to hybrid and discrete event systems. Application areas include automotive and aerospace control systems, network control, biological systems, multiagent and cooperative control, robotics, reinforcement learning, sensor networks, control of cyber-physical and energy-related systems, and control of computing systems.
sort pith recommended most recent
One scalar decides delay-independent stability
Covariance radius eliminates auxiliary matrices from Riccati stability condition, matching Lyapunov's theorem.
New transversality rules when boundary sets are orbits
The result fills a gap for three-body trajectory problems, where endpoint orbits resist algebraic equations.
· “Transversality Conditions for Boundary Constraints Defined by Differential Equations”
One principle reframes why systems stay themselves while changing
Structure-Behavior Coalescence says system identity arises from mutual determination, not from matching separate descriptions
· “Structure-Behavior Coalescence and the Limits of Traditional Systems Theory”
Independent learning in discounted games hits a quasi-polynomial barrier
Under ETH for PPAD, polynomial CCE computation is impossible; first radically uncoupled algorithm matches the hardness bound.
· “Independent Reinforcement Learning in Discounted Markov Games”
Interval reachability ensures safe sampling MPC without pre-computation
Reduces crashes by over 99% in simulation and runs at 60 Hz on a real racecar.
· “Online, Reachability-Aware, Sampling-Based Motion Planning”
Probabilistic method cuts hosting capacity time from hours to seconds
New approach matches Monte Carlo accuracy within 3% while enabling 15-minute grid management cycles.
· “Efficient Quantile-Resolved Hosting Capacity Assessment on Nodal Level for Low-Voltage Grids”
Four labels say when time data survives a viewpoint switch
A sheaf framework classifies persistent versus cumulative views and extends to tree-width and multi-agent networks.
· “Time-Varying Data as Sheaves: an Invitation to Narratives”
99.97% fault and lightning detection for VSC-MTDC grids
Deep learning with S-transform imaging and hybrid attention beats threshold-based methods and generalizes without retraining.
SALT matches EMT accuracy at power-flow speed
The method derives exact frequency-domain equivalents of dynamic device models, revealing that traditional power flow misses up to 75% of…
· “Unifying Power Flow and Electromagnetic Transient Modeling”
High operating costs kill SMR profitability
Even with zero construction cost, most SMRs cannot pay back within their lifetime due to hourly losses from high marginal costs.
· “The consequences of high SMR operating costs in electricity markets”
Asynchronous MPC guaranteed stable and bounded under model mismatch
A competitive ratio ties prediction horizon, inter-execution steps, and mismatch level to stability and performance.
· “Asynchronous Model Predictive Control Under Model Mismatch: Stability and Performance Guarantees”
Frank–Wolfe solves input design with O(1/M) convergence
Each linear subproblem reduces to a budget-constrained LQ problem, solved by Riccati recursion and bisection.
NOFRF estimation goes open source with NL-FRA
MATLAB package offers data-driven least-squares analysis of nonlinear system dynamics directly from input-output data.
· “NL-FRA: A MATLAB package for nonlinear frequency response analysis”
MPC that models hydraulic coupling cuts heating discomfort by 27%
Dynamic valve and temperature control balances heat across rooms without extra energy use
Risk-seeking learners lock onto the high-payoff equilibrium
In coordination games, one risk parameter flips long-run selection between the efficient and the safe outcome.
· “Entropic Risk-Sensitive Evolutionary Learning and Equilibrium Selection in Coordination Games”
Two-stage method kills shedding
Model-based RL uses eight surface pressure sensors to stabilize the unstable cylinder wake equilibrium at Re=100, no full-state needed.
Voltage behaves globally at the 1-second timescale
PMU data from five Irish buses shows voltage synchronizes across the grid at 1-second lags, overturning a textbook assumption.
· “On the Local and Global Nature of Frequency and Voltage Dynamics”
Partial smoothing sharpens descent in nonsmooth bilevel problems
Bi-ZOL smooths only the response Jacobian, achieving O(δ) bias for C^{1,1} responses and exact Clarke stationarity for piecewise affine…
· “Bi-ZOL: Bilevel Zeroth-Order Learning with Nonsmooth Responses”
Koopman controller balances cells as well as nonlinear MPC at 1/9 the compute time
A convex quadratic program replaces nonlinear optimization for simultaneous SoC and temperature balancing in series-connected cells.
Manifold sampling solves hard-constrained trajectory problems
Constraint satisfaction and global exploration combined via Riemannian Adjoint Schrödinger Bridge sampler.
· “Trajectory Optimization via Schr\"{o}dinger Bridge Sampling”
Timeless geometric control enforces temporal logic without clocks
Finite-time level-set inversion maps temporal windows to spatial boundaries, surviving clock snaps and delays.
· “Sound Compilation of Weighted Event Signal Temporal Logic to Timeless Geometric Control”
Config-constrained tube MPC for nonlinear DC systems
Combines DC bounds and fixed-structure polytopes to keep optimization convex while allowing shape to adapt.
· “Configuration-Constrained Tube MPC for Difference-of-Convex Nonlinear Systems”
Measured solar-inverter exposure can crater Oahu grid voltage
A measurement-driven attack simulation shows 41.1 MW of publicly reachable solar DERs can drive half of buses below safe limits.
Transient-aware droop-free control cuts peak cost 10% in weak grids
New method uses dynamic line and load models to improve transient response without altering steady-state behavior.
· “Distributed Droop-Free Control of Grid-Forming Inverters with Dynamic Line and Load Modeling”
Simplex-filter battery control hits optimal √T regret
The controller keeps every step feasible and trails the best hindsight battery policy by only √(T log log T).
· “Online Constrained Control of Storage Systems via Simplex Disturbance-Action Policies”
One shared policy nearly matches an omniscient charging planner
No dispatcher, no reservations: vans trained on four cities run 98.6% of shifts on time across twenty.
· “Emergent Charging Coordination in Electric Delivery Fleets”
One frozen encoder maps any grid from 36 to 70,000 buses
A self-supervised graph encoder reused for power flow, reactive control, operating-state generation, and stability assessment on unseen…
· “A unified power-grid representation for reuse across network structures and computational tasks”
US home heating equipment is 2.75 times oversized
Data from 74 houses show systematic oversizing across climate zones, including the industry standard Manual J.
· “Data-driven estimation of design heating loads for HVAC equipment sizing”
Decentralized shield lets robots avoid collisions without talking
Predictive safety layer and a deadlock-breaking protocol allow pre-trained multi-agent policies to adapt safely to unseen obstacles.
· “Decentralized Safe Multi-Agent Reinforcement Learning via Predictive Shielding”
Spectral Koopman-Hopf tames adversarial reachability
Nonlinear reachable sets computed without state grids by lifting to Koopman coordinates.
· “Spectral Koopman-Hopf Formula for Reachability with Adversary”
Measure how alike two nonlinear systems are
Lift trajectories into observable space, compare subspaces: coordinate changes vanish, real parameter changes register.
· “A Data-Driven Koopman-Behavioral Distance for Nonlinear Dynamical Systems”
Long-duration storage cuts grid load shedding by 83%
A hybrid battery-LDES fleet also stays feasible through every tested transmission outage, where batteries alone fail.
Robotaxi routing can improve both monitoring and speed
Simulation shows a win-win when traffic coverage is added to route planning with the right weighting.
· “RoboSense: Leveraging Robotaxi Fleets as Drive-by Sensors for Urban Traffic Monitoring”
Serving an AI request often happens in a different region from the user
GreenPassport computes request-level carbon for cross-border AI inference with provenance-tiered labels, and its calibrated energy…
· “GreenPassport: Request-Level Carbon Accounting for Cross-Border AI Inference”
Möbius trick gives moment matching for descriptor systems with pole-zero control
Unified framework handles proper and improper systems, places poles and zeros in reduced models, for both continuous and discrete time.
· “Moment Matching for Descriptor Systems: A M\"obius Mapping Approach”
Local energy function judges stability of mixed inverter grids
A sufficient condition for transient stability after large disturbances, with small conservativeness, tested on IEEE 14-bus.
Kh-budget MPC tracks user's violation budget in real buildings
A control framework that adapts temperature bounds to keep realized Kh violations near the budget, tested in simulation and an occupied…
Bayes beats EB at its own game: same risk
A new construction gives the EB estimator's asymptotic accuracy while reducing sensitivity to hyperparameter errors from linear to…
Fine-tuned AI beats custom models for short-term grid load forecasts
Zero-shot Chronos-2 lags; full fine-tuning and cyclical calendar encoding yield best short-horizon results.
· “Assessing Covariate-Informed Grid Load Forecasting with a Time-Series Foundation Model”
Graph transformer beats seven models on mobile demand forecast
TD-STGT predicts grid-level traffic changes using crowdsourced data, achieving a 5.7% RMSE improvement.
· “TD-STGT: A Spatio-Temporal Graph Transformer for Mobile Traffic Demand Forecasting”
State-machine control coordinates multiplexed power converters
Two approaches – Off-Load and Hot-Swap – manage reconfiguration while minimizing voltage disturbances.
· “A Control Philosophy for Multiplexed Power Converters in Active Distribution Systems”
One identity settles how IoT sensors should batch reports
Long-run inference value splits exactly into each report's statistical gain times the chance it arrives by the deadline.
· “Assessing Fixed-Batch Reporting for Deadline-constrained Inference in Intermittently Powered IoT”
Drone plus camera gimbal exactly mimics fixed-wing flight
A control proof and simulation show a camera drone can replace a costly fixed-wing plane for training and testing.
· “Dynamic System Emulation: Fixed Wing Dynamics on a Multicopter”
A single likelihood fit learns IMU drift and sensor noise
Sparse likelihood training turns noisy vision poses into inertial preintegration that survives lighting outages.
· “PLATO: Preintegration Learning from Accurate Trajectory Observations for Neural Inertial Odometry”
Balance scope against age to find the best info radius
Closed-form rules pick when fresh-local data beats a delayed global view and how wide a snapshot should reach.
Backup CBF certifies safety against hidden obstacles
OcclusionCBF extends backup control barrier functions to predicted occupancy, ensuring recursive feasibility and collision avoidance in…
· “OcclusionCBF: Backup Control Barrier Functions for Safe Navigation Among Hidden Dynamic Obstacles”
Partial physics + weak data yields control direction: GDC
Learned score from weak data provides a local value-gradient for online control on a known-feasible scaffold.
· “Geometric Distributional Control: Learning Progress with Partial Structural Knowledge”
Embodied AI safety: rethinking robot safety beyond collisions
A taxonomy of contextual hazards, misalignment, and adversarial misuse, plus a full-stack research agenda.
Verification of unknown stochastic systems improves at a proven rate
Only observations are needed: satisfaction probabilities converge to the truth at a proven rate.
· “Non-parametric Formal Synthesis of Unknown Stochastic Systems: Asymptotic Convergence Guarantees”
First provable rate guaranteed for H∞ policy search
A proximal bundle method keeps every iterate stabilizing and reaches (η, ε)-stationarity in O(max{η^{-4}, ε^{-2}}) steps.
· “Weak Convexity and Proximal Bundle Methods for Nonsmooth Policy Optimization in Robust Control”
This paper proves that for dynamic output-feedback H∞ control
For dynamic output-feedback H∞ policy search, epsilon-Clarke stationarity on the identity-gauge exact slice yields O(epsilon)…
· “Algorithmic Optimality Guarantees for Nonsmooth H_infty Output-Feedback Policy Search”
Cost-induced fixed-time HJI yields uniform settling bound under FDI attacks
Two-power state penalty and replay-based critic learning guarantee convergence independent of initial conditions for nonlinear multi-agent…
Hybrid system tames repeated impacts with permanent dent
By confining switching to a separate contact interface, the model keeps rigid body equations simple while capturing residual deformation.
All observable quantities of rational ODE models now computable
New algorithm uses input-output equations and Lie derivatives to produce compact, interpretable generators
· “Observable functions of rational ODE models and how to find them”
One controller retunes itself as operator goals change
Simulator tests show one deployment swaps objectives, tighter loads, and new constraints by rewriting its control loss.
· “Multimodal Large Language Model-guided Constrained Optimization for RAN Intelligent Control”
GPU simulator brings ground effect to fixed-wing RL at 19.5M steps/s
Open-source FALCON-S unifies semi-empirical aerodynamics, 6DoF physics, and GPU training for systematic controller benchmarking.
· “FALCON-S: Fixed-wing ground-effect Aerodynamics Simulator and Flight Control Learning Suite”
Two-stage backpressure routing unifies interface choice and packet forwarding
By modeling each interface as a virtual subnode, the method jointly assigns packets to modalities and routes them across the network.
· “Biased Backpressure Routing for Multihop Wireless Networks with Heterogeneous Interfaces”
Uplink speed for brain implants is about bandwidth
Review of wireless modalities shows that narrowband links hit a pJ/bit wall; UWB and BCC break through to sub-10 pJ/bit.
· “Bidirectional Wireless Communication for Weakly Coupled Implantable Brain-Computer Interfaces”
Owner-by-owner modelling shifts Germany's heat transition
Agent-by-agent results show owner type and income steer subsidies, grid hotspots, and peak loads that aggregate models miss.
AI labs face same drift that killed Challenger and Boeing
Four mechanisms from past technological disasters are already operating in frontier AI development; safety processes alone are not enough.
Timing watermark catches every robot command hijack in live tests
A tiny Gaussian timing signature catches stealth rewrites that volume-based filters miss.
LLM agents cut urban bottleneck delay 9.18%—with simulator approval
Feedback-guided search finds feasible fixes in 86.7% of cases while simulation, not the LLM, picks the final plan.
141 ms failover cuts jamming-driven robot path error by 78.9%
Hot-standby radios plus spectral detection stop physical-layer jamming from cascading into robot control failure.
Headroom-aware MPC eliminates saturation in PV microgrids
SHCMPC uses hard constraints; SAMPC uses adaptive penalties for lower computational cost.
No LLM is ready for autonomous HVAC control
Systematic analysis of 66 studies shows LLMs are defensible only for bounded semantic and workflow layers, not controllers.
Choosing the right question cuts data-assimilation error 79%
Picking both location and query threshold from the forecast ensemble beats fixed thresholds at equal budgets.
· “Forecast-Ensemble-Based Active Binary-Threshold Query Design for Interval Data Assimilation”
Factorization preserves stability in context-aware control
A structured decomposition of performance-boosting operators keeps L_p stability guarantees when incorporating arbitrary contextual signals.
· “Context-Enriched Performance Boosting via Operator Decomposition”
This paper introduces a new pricing mechanism for electricity markets that uses a causal…
Restricting future dispatch to a shared causal linear policy gives a single locational marginal price with standard energy/congestion…
· “Locational Marginal Pricing for Adaptive Robust Look-Ahead Dispatch with Casual Affine Recourse”
Dust on the route now predicts compressor efficiency loss
A four-layer chain from airport particle dose to nine-stage efficiency loss matches in-service compressor deterioration milestones.
Edge walks match GNNs on grid failures
Random-walk fingerprints weighted by physical line states achieve 98–99% balanced accuracy and beat neural models on failure detection.
· “Physics-Aware Random Walk Fingerprints for Scalable Power Grid Graph Classification”
Constrained Stackelberg games reduce to a single quadratic program
A provable equivalence turns the nested leader-follower problem into one solvable complementarity quadratic program.
Policy iteration tames domain-random LQR
Guaranteed stability and linear convergence for uncertain linear systems.
· “Policy Iteration for Domain Randomized Linear Quadratic Systems”
One integral identity unlocks the matrix exponential derivative
From Taylor series to the Schur–Poincaré adjoint formula, plus a proof the Fréchet derivative exists.
This paper shows that when an LLM tool agent starts in violation
A product order over branch-level violation support and severity is shown to be the necessary representation for both runtime admission…
· “SiLR: Structure-Preserving Admission and Process Reward for LLM Tool Agents”
Thermal digital twin spots impact damage in space habitat
Bayesian inference with an adaptive model pinpoints where and when micro-meteorites hit, even with limited sensors.
· “Bayesian thermal digital twin for a space habitat subjected to an impact event”
LLM + HPC agent solves power flow in under 20 steps
AgentiGrid tests four LLMs on a 200-bus grid; parallel computing cuts iterations but reliability still varies.
· “Integrating Agentic Artificial Intelligence with High-Performance Computing for Grid Planning”
Single-pass learned control keeps OTFS-NOMA sum-rate within 10 percent
A single forward pass per coherence interval keeps sum-rate within about 10 percent across a 128-fold speed range.
Lumped battery models overstate cell health by up to 31 percent
A 12.5 h window resolves all 324 cells; average module models miss the weakest cells that set usable capacity.
First proof that millimeter waves ride along walls
Grazing-incidence peak on drywall, blocked by absorber touch, confirms a new surface-guided path for mmWave signals.
