SO-101 Policy Arena

MolmoAct2 and π0.5 tie on most tasks, usually because neither got past reaching. Both beat GR00T clearly.

Each bar is the mean of 20 rollouts. In dim light the wrist camera is close to black.

Each dot is the mean of 6 rollouts. Tasks are sorted by the best model's score.

Most points come from reaching the object. Grasps were rare, and no model completed a task.

Each setup was run with only one pair, so a setup's numbers only compare those two models.

Protocol and caveats

• Blind A/B. Both models in a pair ran back-to-back from the same object placement. The operator scored each rollout without knowing the model, and the models were revealed after each scene.
• Same execution for all. All models start from the same ready pose. Each predicted action chunk runs open-loop at 30 Hz before the next query, as in the MolmoAct2 paper. Per-tick joint moves are capped at 15° for safety. A rollout lasts at most 60 s.
• Native inputs. Each model gets the front RealSense and the wrist camera at its own input size:
  • MolmoAct2: 378², 30-step chunks
  • π0.5: 224² letterbox, 50-step chunks
  • GR00T: 256², 40-step chunks
  Models ran on a remote A6000, with a median round-trip of about 750 ms (about 360 ms for GR00T).
• GR00T. The processor revision the checkpoint pins is not public, so we used upstream LeRobot. On its own training data GR00T's base turns in the right direction, but it barely beats holding still (mean error 5.65° vs 1.85° for MolmoAct2 on battery_bin_center). The checkpoint authors confirmed the setup.
• Setup is confounded with pair. Lighting and task breakdowns are balanced across pairs. Setup breakdowns are not.
• Two early scenes were relabeled. "Normal table" and "white mat" were recorded before the lighting axis existed and were relabeled "very bright", which matches the conditions they were recorded in.
• Scale. One arm, one operator, one room.

Data