Midhun Xavier

Midhun Xavier

AI Research Engineer @ Aerobase | Neural Combinatorial Optimization & Reinforcement Learning

About me

I am an AI Research Engineer at Aerobase specializing in Neural Combinatorial Optimization (NCO) and Reinforcement Learning (RL). My work bridges theoretical research on learning-based optimization with its applied deployment to complex real-world engineering problems — using deep learning to generate and optimize paths, plan under constraints, and make sequential decisions that improve quality, efficiency, and reliability.

My Expertise

  • Path & Toolpath Optimization – Generating and optimizing robot and machine paths by formulating path planning as a combinatorial optimization problem, sequencing segments to reduce travel, balance load, and improve throughput and output quality.

  • Reinforcement Learning for Quality & Defect Minimization – Training RL agents to minimize defects and improve process outcomes by learning from multiphysics simulation feedback, including finite element (FEM) and computational fluid dynamics (CFD) models.

  • Simulation-in-the-Loop Optimization – Coupling learned policies with physics-based simulators to close the loop between decisions, control parameters, and predicted outcomes.

  • Learning-Based Solvers – Designing neural solvers for classic combinatorial problems (TSP, CVRP, scheduling, bin packing, assignment) that construct and refine solutions directly from data, complementing or replacing hand-crafted heuristics.

  • Attention Models & Architectures – Building Transformer-based encoder–decoder solvers, Pointer Networks, and graph neural networks that scale to large problem instances.

  • Policy Optimization & Reinforcement Learning – Training solvers with policy-gradient and actor–critic methods and modern algorithms such as GRPO, PKPO, RSPO, and PPO, including REINFORCE with greedy and rollout baselines.

  • Pass@K & Max@K Objectives – Optimizing and evaluating models under best-of-K objectives, aligning the training objective with the metric that matters at inference time.

  • Learned Heuristics & Generalization – Developing constructive and improvement (local-search) policies, neural large-neighborhood search, and improving out-of-distribution generalization across instance sizes and distributions.

With a strong foundation in machine learning, reinforcement learning, and optimization, I build neural solvers and RL systems that are efficient, scalable, and grounded in real-world engineering problems.