Sagar Verma

Micropilot. Saarbrücken, Germany.

sagar.jpg

I work on the reliability of learned models in the physical world: build the system, find where it breaks out of distribution, then work the problem behind the break.

My Ph.D. (CentraleSupélec / Paris-Saclay, with Jean-Christophe Pesquet) was on making neural networks safe, reliable, and efficient inside physical systems — certified robustness (Shrink & Cert, CertViT), sparsification by subdifferential inclusion (ICML 2021), and non-linear control with Schneider Electric. Those guarantees do not survive deployed model scale, so my current focus is post-training evaluation and verification of learned policies: searching for the states where a policy is wrong instead of averaging over a held-out set that hides them.

On the application side I lead VLM and VLA work for manufacturing shop floors at Almetra, and low-level nonlinear control for dexterous manipulation and muscle-tendon systems at Micropilotemg2tendon (RSS 2025) and the MJX backend for MyoSuite with Vittorio Caggiano (FAIR, Meta).

Before: CTO and co-founder at Granular AI; IIIT Delhi with Chetan Arora; Watson AI, IBM Research.

Research interests

  • Safety, verification, and post-training evaluation of learned policies
  • Certified robustness and its limits at scale
  • Vision-language and vision-language-action models
  • Dexterous manipulation and low-level control
  • Reinforcement learning and large-scale parallel simulation

Open data, code, and models

  • emg2tendon (RSS 2025) — sEMG to MyoHand tendon control. Project page · Code · Pretrained models · Inverse-dynamics tool
  • MyoSuite MJX — GPU-accelerated musculoskeletal simulation. Code
  • Shrink & Cert (AdvML Frontiers, ICML 2023) — certified training by bi-level optimization. Code
  • CertViT (AdvML Frontiers, ICML 2023) — certification of pre-trained vision transformers. Code
  • robust-motor (NeurIPS 2022 RobustSeq) — robustness of neural models in a physical control loop. Code