Sagar Verma

I am a Staff Computer Vision Engineer at Almetra, where I lead a team building Vision-Language Models (VLMs) for manufacturing plant analysis and, together with our robotics team, extending them into Vision-Language-Action (VLA) models for robotic manipulation — bridging perception to control. My work spans multimodal foundation models, edge deployment, and learning-based control for dexterous, anthropomorphic robots.

Previously, I was the Chief Technology Officer and Co-Founder at Granular AI, where I led the development of GeoEngine, the world’s first Geospatial MLOps platform. Our go-to-market product, Inspect.Properties, served 900+ construction, insurance, and adjuster companies, providing automated roofing, property inspections, and claims documentation using satellite, aerial, and drone imagery. Granular was part of the Techstars Boston 2020 cohort and was one of 8 startups selected for the inaugural NGA Accelerator, engaging with geospatial/GEOINT stakeholders in the Defense and IC communities. GeoEngine was also a US Air Force Research Lab subprime contract, so technically, I’m a defense contractor (cue the Top Gun soundtrack).

Through Micropilot, I work on learning-based low-level control for dexterous, anthropomorphic hands — mapping surface electromyography (sEMG) to tendon control forces for musculoskeletal hand actuation (RSS 2025, RAS 2025). I contribute to MyoSuite in collaboration with its creator Vittorio Caggiano (FAIR, Meta), building its MJX (MuJoCo-XLA / JAX) backend for GPU-accelerated, massively parallel simulation and large-scale RL of dexterous control.

Background

I did my Ph.D. in 2022 from CentraleSupélec, Université Paris-Saclay, under Dr. Jean-Christophe Pesquet, focusing on optimization methods for pruning neural networks for edge deployment. As part of my Ph.D., I worked on modeling the dynamics of heavy electrical motors using neural networks, guided by Dr. Marc Castella, Dr. Nicolas Henwood, and Dr. Al-Kassem Jebai. I was funded by ANRT CIFRE and Schneider Electric.

Before that, I was a Research Fellow at IIIT Delhi (2017-2018), where I worked with Dr. Chetan Arora on egocentric video understanding. I was also part of Watson AI team at IBM Research (2017), developing NLQ systems for banking applications.

Current Research Interests

• Vision-Language & Vision-Language-Action (VLA) Models
• Robotics & Embodied AI / Dexterous Manipulation
• Simulation for Robot Learning (MuJoCo / MJX)
• Reinforcement Learning & Learning-Based Control
• Computer Vision for Manufacturing