Paper 2606.17046
Geometric Action Model for Robot Policy Learning
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 1
- GitHub
- 173 stars
01 In brief
Summary
The paper introduces the Geometric Action Model (GAM), a language-conditioned manipulation policy that repurposes a pretrained geometric foundation model (GFM) as a shared substrate for perception, temporal prediction, and action decoding.
GAM splits the GFM at an intermediate layer: shallow layers encode observations, a causal future predictor inserted at the split forecasts future latent tokens conditioned on language, proprioception, and action history, and the remaining deep GFM blocks decode both future geometry and action chunks.
This design integrates 3D geometric priors directly into the policy, addressing limitations of 2D-based vision-language-action models (VLAs) and video world-action models (WAMs).
GAM is trained end-to-end with losses for action regression, future-feature alignment, and future-depth supervision.
Experiments on LIBERO, LIBERO-Plus, RoboCasa-Kitchen, and real-world tasks show GAM matches or exceeds baselines in success rate while being faster (6.9 ms inference, up to 55x faster than Cosmos Policy) and lighter (1.4B parameters).
It shows strong robustness to camera perturbations, improving by 9.7 percentage points over baselines on LIBERO-Plus.
Ablations confirm the importance of pretraining, split-layer selection, and future-prediction losses.
Limitations include bounded language reasoning from the frozen text encoder, suggesting future integration of large language models.
02 From the paper
Abstract
Generalist robot policies must follow user instructions while reasoning about how objects, cameras, and robot actions interact in the 3D physical world. Recent vision-language-action models (VLAs) and video world-action models (WAMs) inherit strong semantic or temporal priors from large-scale foundation models, but they still operate primarily on 2D image frames or 2D-derived latent spaces, leaving implicit the 3D geometry required for contact-rich manipulation. We propose the Geometric Action Model (GAM), a language-conditioned manipulation policy that directly repurposes a pretrained geometric foundation model (GFM) as a shared substrate for perception, temporal prediction, and action decoding. GAM splits the GFM at an intermediate layer: the shallow layers serve as an observation encoder, and a causal future predictor inserted at the split layer forecasts future latent tokens conditioned on language, proprioception, and action history. The predicted future tokens are then routed through the remaining GFM blocks for feature propagation and decoding, allowing a single backbone to produce both future geometry and actions. This design equips the GFM with language-conditioned temporal world modeling through minimal architectural modification while preserving its rich geometric priors. Across a broad suite of simulation and real-robot manipulation benchmarks, GAM is more accurate, more robust, faster, and lighter than current foundation-model-scale baselines.