Oliver Groth

now training humanoids at Vesoma

AI researcher, robot trainer & video game nerd.

I study how machines can learn from their own experience: through curiosity, exploration and a feel for physics. I am trying to understand embodied intelligence by building it. Today that means creating self-learning humanoids at Vesoma. Before that, I worked on AI for robotics at DeepMind and wrote my PhD thesis at Oxford on the acquisition of physical intuition by machines.

selected research ↓ the path so far

About

I work on humanoid robot research at Vesoma. Before that I was a Research Scientist at DeepMind, working on exploration for RL, continuous control and vision-language-action models (VLAs) in the team of Martin Riedmiller.

My PhD at Oxford, with Ingmar Posner and Andrea Vedaldi, asked how robots can acquire physical intuition and put it to use. In my TEDx talk I explain why that might teach a robot to play Jenga.

Earlier stops: a Diplom in Computer Science at TU Dresden with Carsten Rother, and the Stanford Vision Lab with Fei-Fei Li, where I worked on VisualGenome.

The path so far

full cv →
  1. TU Dresden logo

    2011 – 2017

    TU Dresden

    Diplom Computer Science

  2. Stanford logo

    2015 – 2016

    Stanford

    Computer vision research

  3. Oxford logo

    2017 – 2022

    Oxford

    PhD in Robotics

  4. DeepMind logo

    JAN 2022 – JUL 2026

    DeepMind

    AI research

  5. Vesoma logo

    ▶ SINCE AUG 2026

    Vesoma

    Humanoid robot research

Selected research

all publications →

arXiv · 2026

EXIMO

A VLM plans for a VLA to explore new tasks, then imitation and RL finetune it.

Blog · 2026

Gemini Robotics 2

Whole-body control, fine dexterity and teamwork for robots of every shape.

ICML · 2024

Offline Actor-Critic at Scale

Offline RL scales to large transformers and beats behaviour cloning on 132 control tasks.

arXiv · 2021

Is Curiosity All You Need?

Which useful behaviours emerge when a robot explores purely out of curiosity.

ECCV · 2018

ShapeStacks

Learning vision-based physical intuition for stacking objects.

IJCV · 2017

Visual Genome

Dense annotations connecting images and language through objects and relationships.

Play

AFTER HOURS