Vassar Robotics · Reproducible evaluation
See every rollout, not just the score.
Explore MolmoAct 2 across LIBERO tasks with synchronized model-facing cameras, exact task definitions, simulator success criteria, and content-bound provenance.
Paper evaluation
Four-suite reproduction
The paper score includes only LIBERO-10, Goal, Object, and Spatial.
- Paper success rate
- 97.2%
- 1,944 of 2,000 completed rollouts
- Completed tasks
- 40
- Four paper suites
- Completed rollouts
- 2,000
- Fixed-seed episodes
- Camera videos
- 4,000
- Agent + eye-in-hand
Pinned standard-checkpoint paper reference: 97.2%. This reference is not blended with LIBERO-90.
Supplemental evaluation
LIBERO-90 stays separate
LIBERO-90 adds breadth across ninety tasks but was not a MolmoAct 2 paper benchmark. Its result is intentionally reported as a distinct supplemental scope.
- Success rate
- 31.5%
- 1,419 successes
- Completed rollouts
- 4,500
- 90 completed tasks
Benchmark suites
Explore by suite
Open a suite to compare task-level rates and inspect every rollout.
LIBERO-10
Ten long-horizon tasks from the MolmoAct 2 paper evaluation protocol.
LIBERO-Goal
Tasks that vary the target goal while retaining familiar scenes and objects.
LIBERO-Object
Tasks centered on manipulating different object categories.
LIBERO-Spatial
Tasks that test spatial relationships and placement precision.
LIBERO-90
A 90-task supplemental evaluation, reported separately from the paper score.
What you are seeing
The model’s actual visual inputs
Each rollout records both 256×256 observations after LIBERO’s image processor and before the policy processor. The H.264 streams are silent and synchronized by control step.
observation.images.imageobservation.images.image2