Skip to main content
Publication / 2026.06

Geometric Action Model for Robot Policy Learning

A language-conditioned robot policy that repurposes a pretrained geometric foundation model for perception, future geometry prediction, and action decoding.

Role Equal Contribution
Venue 3DWM @ ECCV 2026
Date June 2026
Authors Jisang Han*, Seonghu Jeon*, Jaewoo Jung, René Zurbrügg, Honggyu An, Tifanny Portela, Marco Hutter, Marc Pollefeys, Seungryong Kim†, Sunghwan Hong†
Paper Project Page arXiv Hugging Face PDF Code
Geometric Action Model for Robot Policy Learning
Contents
  1. Abstract
  2. Notes
  3. Links

Abstract

Geometric Action Model (GAM) is a language-conditioned manipulation policy that directly reuses a pretrained geometric foundation model as the shared backbone for perception, temporal prediction, and action decoding. GAM splits the geometric model at an intermediate layer: early blocks encode observations, a causal future predictor forecasts latent tokens conditioned on language and robot state history, and the remaining blocks propagate those tokens into future geometry and actions.

Across simulation and real-robot manipulation benchmarks, GAM improves accuracy, robustness, inference speed, and model efficiency compared with foundation-model-scale baselines.

Notes

  • * Equal contribution
  • † Corresponding author
#3D Vision #Geometric Foundation Models #Robot Learning #Visuomotor Control