Research

I develop safe, interpretable, and adaptive AI systems for real-world cyber-physical environments that must operate under uncertainty, strict physical constraints, and adversarial conditions. My work sits at the intersection of machine learning, control theory, optimization, and AI safety, with a core focus on building models that are robust, predictable, and deployable at scale.

Vision

Full Publications β†’

I aim to create AI agents that generalize across environments, learn from imperfect or unreliable data, and remain stable under distribution shifts. These capabilities are essential both for scientific progress and for building production-grade intelligent systems used in industry.

My research centers around five pillars:

  • Safe & Trustworthy Reinforcement Learning: Designing agents that remain reliable under sensor noise, hardware faults, non-stationarity, and adversarial perturbations. This includes constraint-aware learning, certified robustness, and safe exploration in safety-critical settings.
  • Physics-informed Deep Reinforcement Learning (DRL): Embedding physical laws, invariants, and feasibility constraints directly into model architectures and learning objectives to improve convergence, interpretability, and real-world deployability.
  • Probabilistic & Bayesian Modeling: Quantifying epistemic and aleatoric uncertainty for risk-aware planning in partially observable, high-stakes environments using Bayesian neural networks, uncertainty-aware RL, and probabilistic inference.
  • Large Language Models (LLMs) for autonomous reasoning: Leveraging LLMs to support high-level planning, explainable decision pipelines, natural-language supervision, and adaptive control. This enables human-AI collaboration and interpretable reasoning in complex systems.
  • Vision-based simulation environments: Using platforms such as CARLA, CityLearn, AirSim, and OpenDSS to train agents in multimodal, visually rich environments, supporting robust perception-control integration and sim-to-real transfer.

Mission

By integrating physics-guided structure, probabilistic reasoning, and safe reinforcement learning, my mission is to build the next generation of AI systems that are:

  • Reliable, even under uncertainty, noise, and adversarial conditions
  • Generalizable across tasks, scales, and distribution shifts
  • Interpretable to operators, engineers, and decision-makers
  • Deployable in large-scale, real-world cyber-physical environments

My research advances the foundation needed to deploy trustworthy AI across critical infrastructures, autonomous systems, and safety-critical infrastructure domains where reliability and robustness are essential for real-world impact.

Research Focus

These focus areas organize my ongoing and recent projects that bridge fundamental methods and deployable systems.

DRL-based Control
DRL Control DRL for Volt-VAR
Design control agents for voltage regulation and reactive power optimization in smart distribution grids. Actor Critic Physics-Informed Actor-Critic
Embed grid physics and control limits directly into the DRL learning loop for stable and efficient decisions. Sim2Real Transfer Sim-to-Real Transfer
Train agents in simulated OpenDSS environments and deploy them on real-time OPAL-RT setups.
Safe & Trustworthy RL
Safe RL Robust & Stable Learning
Develop agents that ensure system safety, robustness, and interpretability under uncertainty. Uncertainty Quantification Uncertainty-Aware Policies
Quantify epistemic and aleatoric uncertainty in high-stakes, partially observable settings.
Transfer & Meta-Adaptation
Transfer Learning Domain Adaptation
Enable agents to generalize across grids with different topologies, dynamics, and loads. Meta RL Meta-RL for Efficiency
Leverage meta-reasoning to accelerate learning in low-data, high-variance scenarios.
Vision-Simulation Integration
Perception Control Perception-Control Fusion
Use CARLA and AirSim to train end-to-end systems in visual RL tasks with sensors. Multimodal Fusion Multi-modal Representations
Combine visual, state, and contextual features for better decision-making.
LLM-Augmented Decision Systems
LLM-guided control LLM-Guided Control
Translate natural language into actionable policies for real-world environments.

Publications

Journal Papers Total: 2

  1. Arif Hussian, Kundan Kumar, Gelli Ravikumar
    Bayesian-optimized bidirectional long-short-term memory network for wind power forecasting with uncertainty quantification , Electric Power Systems Research, 2026
    Paper Code ️Poster
  2. Kundan Kumar, Gelli Ravikumar
    Physics-based Deep Reinforcement Learning for Grid-Resilient Volt-VAR Control (Under Review), IEEE Transactions on Smart Grid, 2025
    Paper Code Poster

Conference Papers Total: 7

  1. Kundan Kumar, Gelli Ravikumar
    A Multi-Objective Optimization Framework for Carbon-Aware Smart Energy Management , IEEE North American Power Symposium (NAPS), 2025
    Paper Presentation
  2. Kundan Kumar, Kumar Utkarsh, Wang Jiyu, Padullaparti Harsha
    Advanced Semi-Supervised Learning With Uncertainty Estimation for Phase Identification in Distribution Systems , IEEE PES General Meeting, 2025
    Paper Presentation Poster
  3. Kundan Kumar, Gelli Ravikumar
    Transfer Learning Enhanced Deep Reinforcement Learning for Volt-Var Control in Smart Grids , IEEE PES Grid Edge Technologies Conference & Exposition, 2025
    Paper Poster

Ongoing Projects

  • Federated DRL for Cyber-Resilient Volt-VAR Optimization
    Decentralized, communication-efficient control using LSTM-enhanced PPO agents across distributed DERs.

  • One-Shot Policy Transfer with Physics Priors
    Train agents on small topologies and adapt to IEEE 123-bus, 8500-node networks in a few iterations.

  • LLM-Guided Autonomous Planning for Smart Buildings
    Convert user prompts to interpretable control policies using LLMs (OpenAI, Claude) in CityLearn environments.