Geometric Action Model for Robot Policy Learning

The key detail the summary gestures at but doesn't unpack is what 'repurposing a pretrained geometric foundation model' actually means in practice: rather than training a new perception stack, GAM treats depth, surface normals, and spatial structure as first-class inputs to the policy, which changes how the model generalizes across object configurations it hasn't seen before.

GAM belongs to a broader pattern in recent ML research where pretrained foundation models are being redirected toward constrained, structured tasks without full retraining. The 'Exact Posterior Score Estimation for Solving Linear Inverse Problems' paper covered the same week takes a comparable approach in a different domain, showing that a pretrained diffusion model can be repurposed for measurement-constrained inference by exploiting its learned structure rather than fine-tuning it. Both papers are essentially asking the same architectural question from different angles: how much task-specific work can be offloaded to the geometry or statistics already baked into a foundation model? For robotics, the practical stakes are higher because sim-to-real transfer failures are expensive and slow to diagnose.

Watch whether GAM's contact-rich manipulation benchmarks (particularly any dexterous grasping evaluations) hold up when tested on object categories absent from the geometric foundation model's pretraining distribution. If performance degrades sharply there, the approach is more dependent on pretraining coverage than the architectural framing suggests.