MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation

Procedural environment generation and large-scale simulation have shown promise in training robust robotic policies. However, zero-shot transfer of these policies to diverse real-world tasks remains a challenge. We introduce MolmoB0T, a suite of general-purpose manipulation policies trained on a massive dataset of 2.5 million expert trajectories in simulation. We leverage a diverse set of over 10,000 articulated objects and 500 procedurally generated environments to ensure broad coverage of manipulation tasks. We train three policy classes - MolmoBot, a Molmo2-based multi-frame vision-language model with a flow-matching action head; MolmoBot-Pi0, which replicates the pi_0 architecture to enable direct comparison; and MolmoBot-SPOC, a lightweight policy suitable for edge deployment and amenable to RL fine-tuning. We evaluate on two robotic platforms - the Franka FR3 for tabletop manipulation tasks and the Rainbow Robotics RB-Y1 mobile manipulator for door opening, drawer manipulation, cabinet interaction, and mobile pick-and-place. Without any real-world fine-tuning, our policies achieve zero-shot transfer to unseen objects and environments. On tabletop pick-and-place, MolmoBot achieves a success rate of 79.2% in real-world evaluations across 4 settings, outperforming pi_0.5 at 39.2%. Our results demonstrate that procedural environment generation combined with diverse articulated assets can produce robust manipulation policies that generalize broadly to the real world.

ICRA 2026 SDRL Workshop