MolmoSpaces

The Allen Institute for AI's open ecosystem of large-scale indoor simulation scenes and robot evaluation tools.

MolmoSpaces is an open ecosystem released by the Allen Institute for AI (Ai2) in February 2026, for generating data and training and evaluating robot policies at scale. It includes over 230,000 indoor scenes (drawn from iTHOR, ProcTHOR, Holodeck, and others) and 130,000 annotated object models, of which 48,000 are manipulable objects paired with 42 million stable grasp poses. Scenes aren't tied to any one simulator, supporting MuJoCo, ManiSkill, and Isaac. Its companion benchmark, MolmoSpaces-Bench, contains 8 tasks covering tabletop and mobile manipulation, navigation, and cross-room long-horizon tasks; the paper reports that its simulated scores correlate strongly with real-robot scores (R = 0.96), and finds that policies are quite sensitive to instruction wording, initial joint positions, and camera occlusion.

ExampleAi2's MolmoBot procedurally generates 1.8 million simulated trajectories in MolmoSpaces and deploys directly to a Franka FR3 with no real-robot data at all, reaching a real-robot tabletop pick-and-place success rate of 79.2%, compared with 39.2% for π0.5.

Also called

MolmoSpaces-Bench

Related

Allen Institute for AI · Simulation-Based Evaluation · Sim-to-Real Correlation · ProcTHOR (Large-Scale Embodied AI Using Procedural Generation) · Objaverse · Mobile Manipulation

Sources

MolmoSpaces: A Large-Scale Open Ecosystem for Robot Navigation and Manipulation (arXiv 2602.11337)
Ai2 Blog: MolmoSpaces
MolmoBot: Large-Scale Simulation Enables Zero-Shot Manipulation (arXiv 2603.16861)

As of

2026-03