MolmoAct2

MolmoAct2 is AI2's second-generation action reasoning model with an unusually broad Apache-2.0 checkpoint zoo: base, Think (depth-reasoning), DROID (Franka), Bimanual YAM, SO100/101, and LIBERO variants are all confirmed. WorldFoundry ports the official DROID/YAM predict_action server contracts into an in-tree action_trace pipeline with no external checkout dependency.

The recorded paper is arXiv 2605.02881.

Code: allenai/molmoact2.

worldfoundry-eval run \
  molmoact2-droid \
  --pipeline.task-profile interactive-video \
  --json

What it is for

molmoact2-droid is the recorded instruction-to-action line. Start here when you need a WorldFoundry VLA rollout rather than a video generator.

MolmoAct2 builds on the Molmo open VLM line with action-reasoning heads; the Think variant adds depth-based reasoning before acting. Configured DROID, YAM, SO100, LIBERO, and Think-LIBERO profiles expose checkpoint-backed action_trace routes. It is published by Allen Institute for AI (AI2).

Catalog tasks include vla, vla action prediction, vla policy rollout, and robot policy. Catalog aliases include molmo-act2, molmoact, molmoact2-droid, and molmoact2-yam. Also recorded: molmoact2-bimanual-yam, molmoact2-so100-101, molmoact2-libero.

Recorded evaluation entry points include LIBERO (Dedicated LIBERO and Think-LIBERO checkpoints).

Runtime contract

Optional prompt, input_path, interactions, and fps; artifacts action_trace and molmoact2_action_trace.json.

CLI task profiles: interactive-video. WorldFoundry binds MolmoAct2Pipeline.

Start on molmoact2-droid. The family registers 5 catalog variants (cards above): molmoact2-droid (vla.policy_rollout) — Static load only; molmoact2-bimanual-yam (vla.policy_rollout) — Static load only; molmoact2-so100-101 (vla.policy_rollout) — Static load only; molmoact2-libero (vla.policy_rollout) — Static load only; molmoact2-think-libero (vla.policy_rollout) — Static load only. Family status: Verified.

Pick the checkpoint per embodiment: DROID for Franka arms, BimanualYAM for dual-arm YAM, SO100_101 for SO-ARM, LIBERO/Think-LIBERO for simulation. Integration is Static load onlyally verified.

WorldFoundry in-tree MolmoAct2 runtime lives under worldfoundry/synthesis/action_generation/molmoact2 and no longer depends on the external molmoact2 checkout at runtime.

Runnable / integrated: molmoact2-droid, molmoact2-bimanual-yam, molmoact2-so100-101, molmoact2-libero, and molmoact2-think-libero.

Running in WorldFoundry

Launch worldfoundry-eval run molmoact2-droid --pipeline.task-profile interactive-video to execute the policy. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include allenai/MolmoAct2-DROID, allenai/MolmoAct2-BimanualYAM, and allenai/MolmoAct2-SO100_101 among others. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh molmoact2-droid. Add --download to fetch missing public checkpoints.

Recorded licenses include apache-2.0; status: Verified.

Related models

BeingBeyondBeing-H0.5

Open recipe

GigaAIGigaBrain-0

Open recipe

Open-source research release (institution not recorded)OpenVLA

Open recipe

Open-source research release (institution not recorded)RoboFlamingo

Open recipe