Summary
The Geometric Action Model (GAM) repurposes a pretrained Geometric Foundation Model (GFM) as a unified substrate for perception, temporal prediction, and action decoding, bypassing the 2D-only representations that limit standard VLAs and WAMs in contact-rich tasks. The GFM is split at an intermediate layer: shallow layers serve as an observation encoder, and a causal future predictor inserted at the split point forecasts future latent tokens conditioned on language, proprioception, and action history. Actions are decoded from the predicted geometric latents. The result is a language-conditioned manipulation policy that achieves 55× faster inference than foundation-model-scale baselines while improving accuracy and robustness.
Key Contributions
- Introduces the concept of repurposing a geometric foundation model as a robot policy backbone
- Causal future predictor operating in GFM latent space enables implicit 3D world modeling
- Language and proprioception conditioning for instruction-following manipulation
- 55× inference speedup over competing foundation-model baselines with better accuracy
Significance
GAM demonstrates that 3D geometric structure is a high-leverage prior for contact-rich manipulation; the approach bridges geometric perception research and robot learning in a single compact architecture.