GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks
The paper introduces Geometric Anchor Pre-training (GAP) as a method for improving data efficiency in robotic manipulation tasks. GAP aims to enhance the robustness of visuomotor learning by pre-training a spatial adapter on a simulated task before downstream imitation learning. The results demonstrate that GAP outperforms existing methods under conditions of limited data and domain shifts.
- ▪GAP regularizes a spatial adapter to produce stable geometric anchors for few-shot policy learning.
- ▪The method was evaluated on RoboMimic and ManiSkill with severe data scarcity, achieving notable success rates.
- ▪GAP is lightweight and can be reused across different environments and manipulation skills.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Robotics arXiv:2605.15836 (cs) [Submitted on 15 May 2026] Title:GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks Authors:Davide Buoso, Andrea Protopapa, Stefano Di Carlo, Francesca Pistilli, Giuseppe Averta View a PDF of the paper titled GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks, by Davide Buoso and 4 other authors View PDF HTML (experimental) Abstract:Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
