GAM: Grounded Action Model Proposes 3D Grounding as Robot Foundation Base

Researchers from Allen AI, the University of Washington and NUS (Jiafei Duan, Ranjay Krishna et al.) propose the Grounded Action Model (GAM), questioning the mainstream assumption that language or video models are the right foundation for robot foundation models. GAM instead builds on pretrained 3D grounding models — "ground first, then act" — reaching 61% on LIBERO-PRO, with the authors reporting that smaller models under this paradigm outperform larger VLAs.

Confirmed

• GAM is a new paradigm for robot foundation models centered on metric-level 3D grounding, training robot policies directly on 3D understanding/detection models rather than on VLMs or video models
• The motivation draws on developmental psychology: per the posts, infants do not master language before understanding the world (Hespos & Spelke on conceptual precursors of language), suggesting grounding may be a more fundamental cognitive ability than language
• Both existing routes have shortcomings: the pretrained backbones of VLAs and world-action models (WAM) do not explicitly require metric-level object grounding, which is only learned implicitly from robot demonstrations
• Per author @DJiafei, smaller models in this paradigm outperform larger VLA models; the official NUS post reports notably higher manipulation success rates
• GAM achieves 61% on the LIBERO-PRO benchmark and supports language-based interfaces

Why it matters

• The work directly challenges the dominant LLM/video-backbone VLA route and offers an alternative foundation for robot foundation models
• If the "ground first, then act" paradigm holds, robot policy learning may no longer depend on massive language/video pretraining, achieving stronger spatial manipulation with smaller models

2026-09-22 ~ 2026-09-23 · 6 related posts

• [source] NUS's Grounded Action Model Tops Robot Manipulation Benchmarks with 3D Grounding — NationalUniversityofSingapore · 2026-09-22
• [source] Grounded Action Model: Building Robot Foundation Models on 3D Grounding, Hits 61% on LIBERO-PRO — DJiafei · 2026-09-22
• Grounded Action Model builds robot foundation models on pretrained 3D grounding instead of VLMs — DJiafei · 2026-09-22
• Grounded Action Model: building robot foundation models on 3D grounding instead of VLAs or video models — DJiafei · 2026-09-23
• GAM: training robot policies from a 3D detector beats larger VLAs, with 1M-image WildDet3D dataset — DJiafei · 2026-09-23

1 near-duplicate retellings: DJiafei