Milan Ganai

I am a third-year PhD student in the Computer Science Department at Stanford University advised by Marco Pavone and Clark Barrett. I received my Bachelor of Science (2020-2023) and Master of Science (2023-2024) in Computer Science at UC San Diego, where I was advised by Sicun Gao and Sylvia Herbert. My interests are in multimodal reasoning and out-of-distribution (OOD) generalization for robotics.

Contact: mganai at cs dot stanford dot edu

Google Scholar  |  LinkedIn

Vision-Language Models are increasingly used as high-level planners for embodied agents, and the default recipe for improving them is to scale test-time compute like chain-of-thought reasoning, bigger models, more memory history. In robotics, that compute is not free: it adds latency, tokens, and FLOPs while delivering uneven, often diminishing returns. DIRECT is a lightweight routing framework that uses multimodal scene context to decide when and where to spend test-time compute, learning a mapping from a task's scene and instruction to the planner best suited to solve it, matching high test-time-compute-level success at a fraction of the planning time.

^➤equal contribution

Embodied Chain-of-Thought (CoT) reasoning has enhanced Vision-Language-Action (VLA) models, but rigid templates over reasoning primitives (objects, plans, affordances) force policies to process irrelevant information that distracts from action-prediction signals. We introduce R&B-EnCoRe, which treats reasoning as a latent variable within importance-weighted variational inference, enabling models to bootstrap embodiment-specific reasoning from internet-scale knowledge through self-supervised refinement — without external rewards, verifiers, or human annotation. Validated across manipulation, legged navigation, and autonomous driving with various VLA architectures with 1B, 4B, 7B, and 30B parameters.

^♫equal contribution

As autonomous systems expand their deployment regions into unstructured, open-world environments, they face potential hazardous Out-of-Distribution (OOD) failure scenarios that differ from their training data. Current methods rely on handcrafted intervention policies, limiting their ability to plan generalizable, safe motions. FORTRESS introduces a novel framework that generates and reasons about semantically safe fallback strategies in real time to prevent OOD failures by bridging open-world, multi-modal reasoning with dynamics-aware planning.

A journal publication surveying the recent literature on scalable Hamilton-Jacobi reachability estimation in reinforcement learning to provide a foundational basis for research into reliability in high-dimensional systems. We review how this technique has been employed to solve challenging tasks like those with dynamic obstacles and lidar-based or RGB image-based observations.

Hamilton-Jacobi reachability estimation for model-free safe RL in deterministic and stochastic environments with safety guarantees and convergence analysis. Tasks include lidar-based observations, dynamic obstacles, and multiple hard and soft constraints.

Learning Lyapunov-like models offline from observation-only, expert data to solve stabilization control tasks online. Deployed in hardware for robustness testing.

Reinforcement Learning to determine XLA compiler optimization pass ordering to reduce GPT-2, BERT, and ResNet graph sizes.

Building boosted decision trees to classify reconstructed particle tracks as merged in high density particle physics environments.

Original code template