Egocentric data for robot learning: what recent work shows

October 10, 2026 · 3 min read · Data collection, Research · By the ARX Infinity team

Egocentric data for robot learning: what recent work shows

Short answer: egocentric data is video and motion recorded from a person's own point of view while they do a task. It can be collected in many more places than robot teleoperation, and recent work uses it at large scale, mostly in combination with robot data. The practical setup is a human-data pipeline for volume plus a few real robot stations for training on robot data and testing policies.

What the recent work shows

  • EgoVerse (2026) is a shared platform and dataset with 1,362 hours of egocentric human demonstrations, from Georgia Tech, Stanford, UC San Diego, ETH Zürich and industry partners. Both of its robot setups use two 6-DoF ARX5 arms (paper).
  • EgoWAM (Georgia Tech, 2026) trains manipulation policies on egocentric video by predicting both the actions and how the scene will change (project).
  • UMI (2024) records demonstrations with handheld grippers instead of a robot. Its deployment code for ARX arms is listed in Stanford's arx5-sdk (UMI).
  • Xiaomi-Robotics-1 (2026) was pre-trained on more than 100,000 hours of embodiment-free UMI data, then post-trained on more than 7,200 hours of in-house robot data (project).
  • SIM1 (2026) turns about 200 human demonstrations into synthetic cloth-manipulation data in a physics-aligned simulator (project).

The pattern across these projects: human data supplies volume and variety, and robot data on real arms grounds the policy in the robot's own body.

Where it helps and where it does not

Helps with Still needs robot data for
Many scenes, objects and people The robot's own kinematics and gripper
Long-tail tasks that are rare in the lab Contact and force behaviour
Collecting where a robot can't go Final evaluation of the policy

A practical setup

  1. Record egocentric demonstrations for breadth. ARX's AC 2, released in February 2026, is ARX's system for egocentric (EGO) data and reinforcement learning (RL).
  2. Keep bimanual robot stations for robot data and testing. The bimanual platform and AC one record leader–follower demonstrations on ARX arms.
  3. Test on the same arms other labs use. EgoVerse's robot setups use ARX5 arms, which makes comparison easier.

Sources

Get a price for your setup.

Price, lead time and shipping options for your country.

Request a quote