GAM Hits 61% on LIBERO-PRO by Making 3D Grounding Explicit

On LIBERO-PRO, a new robot policy called GAM posts a 61% average success rate across 16 perturbation settings, eight points above the 53% for π0.5. The gap widens on real hardware: on a bimanual YAM under visual shift, GAM keeps 17 of 20 successes where π0.5 keeps only 4.

The pitch, in a paper on arxiv, is that current robot foundation models leave metric grounding to emerge from demonstrations rather than requiring it in the backbone. "Manipulation policies must know which objects matter and where they are," the authors write, yet the pretrained backbones behind vision-language-action and world-action models "do not directly require this metric grounding, leaving it to be learned implicitly from robot demonstrations." GAM instead accepts language, point, or box prompts, turns them into an object-centric representation that carries both visual features and metric geometry, and feeds a multi-stream transformer that predicts action chunks.

The benchmark story is not one number. On RoboTwin 2.0, GAM averages 55.3% across 50 tasks against 52.0% for Spatial Forcing, and 47.6% under scene randomization against 30.4% for Abot-M0, with its action policy trained only on clean-scene demonstrations. Composed with a Molmo2 planner on a Franka arm, it reaches 64.7% in-distribution and 49.8% out-of-distribution step completion on long-horizon, memory-dependent tasks.

The paper reports the largest LIBERO-PRO gains "when targets are relocated or newly designated," the setting where explicit grounding should help most. All comparisons are the authors' own runs on their own harness; no independent lab has yet reproduced the bimanual YAM visual-shift result.