Jisang Han, Seonghu Jeon, Jaewoo Jung, Renée Zurbrügg, Honggyu An, Tifanny Portela, Marco Hutter, Marc Pollefeys, Seungryong Kim, Sunghwan Hong
4 min
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.
Sam: I'd have assumed all that 3D reasoning would make it slower. Is there a computational cost?
Alex: You'd expect that, but the research reports the opposite. Because the same backbone handles both perceiving the scene and planning the action—rather than running separate systems—it's substantially faster than approaches that use a technique called diffusion to generate movements. Diffusion models are powerful, but they work by iteratively refining a noisy guess into a clean answer, which takes time. GAM sidesteps that entirely.
Sam: So the speed comes from the fact that understanding the scene and deciding what to do are happening inside the same structure, rather than being two separate, expensive processes?
Alex: That's the key insight. The joint prediction—forecasting the future geometry of the scene at the same time as the robot's own next actions—is what the researchers identify as the central mechanism behind the performance gains. The model stays oriented even when the camera is moved, because it's reasoning about a 3D world, not just a particular camera angle.
Sam: It's a meaningful shift from just training a robot to copy motions it's seen before. This approach is asking the robot to actually understand the space it's operating in.
Alex: That's the core argument of the paper. By integrating geometric awareness directly into how the robot makes decisions, the system no longer has to infer depth and position from flat pixels—that understanding is already there. The evidence the researchers present suggests this is a clear improvement, particularly when conditions change in ways the robot hasn't seen before. Thanks for listening to ResearchPod.