Open-Source Guide: Extracting Robot Actions from Egocentric Video

gui_penedo · x · 2026-08-07

Macrodata Labs published an in-depth technical blog on building an RGB-only pipeline to extract metric 3D hand trajectories from egocentric video, providing supervision for robot actions without specialized hardware.

Evaluation & Challenges

The team evaluated full hand-tracking pipelines on the HOT3D dataset using Action MPJPE. A viable system needs >75% direct-prediction coverage and 15 FPS on an H100.

System & Results

Their final system combines WiLoR detection, HaWoR reconstruction, and VGGT-Omega camera trajectory. It achieves 52.04 mm Action MPJPE and 15.53 FPS, reducing error from 59.12 mm and increasing speed from 3.34 FPS compared to the original HaWoR pipeline. Hand reconstruction was identified as the largest error source, with depth accounting for 43% of the action error.

Related event: Open-Source Pipeline Extracts Robot Actions from Egocentric Video(2 posts)→

Original post →

More from Embodied

Embodied channel →