DEVOURED
AI2 released MolmoAct 2, an open robotics model that nearly doubles Physical Intelligence's proprietary π0.5 success rate on real-world tasks, plus the field's largest open bimanual dataset at 720 hours and full training code.
What: AI2 released MolmoAct 2 on May 5, 2026, an Action Reasoning Model achieving 87.1% success on real-world Franka arm tasks versus Physical Intelligence's π0.5 at 45.2%. Runs in 180ms per action (37x faster than MolmoAct's 6,700ms). Released with MolmoAct 2-Bimanual YAM dataset containing 720+ hours of two-arm training demonstrations. Built on Molmo 2-ER backbone scoring 63.8/100 on embodied reasoning benchmarks ahead of GPT-5 and Gemini 2.5 Pro. Fully open-source: model weights, datasets, training pipeline, and reference hardware setup published.
Why it matters: This demonstrates open robotics foundations can now beat proprietary systems on industry benchmarks while being radically more transparent. Most teams release model weights but withhold training data and code. AI2's full release—including the 720-hour bimanual dataset, training code, and hardware specs—establishes a new openness standard for physical AI where real deployment in unstructured environments requires reproducible foundations the research community can build on.
Takeaway: Download MolmoAct 2, the MolmoAct 2-Bimanual YAM dataset, and reference hardware specs from AI2 if working on robotics manipulation research or real-world deployment.
Deep dive
- MolmoAct 2 built on Molmo 2-ER, trained on ~3M additional embodied-reasoning examples (pointing, detection, spatial reasoning, video QA), scoring 63.8/100 across 13 benchmarks ahead of GPT-5, Gemini 2.5 Pro, Qwen3-VL-8B
- Architecture pairs Molmo 2-ER with action expert using flow matching via KV-cache bridge, includes open MolmoAct 2-FAST tokenizer reimplementation
- Inference: 180ms per action base model, 790ms with adaptive depth reasoning (versus 6,700ms original MolmoAct on H100 in LIBERO benchmark)
- MolmoAct 2-Think variant adds depth perception tokens with adaptive-depth routing that predicts depth only for dynamic scene regions, achieving 17% speedup versus full depth prediction
- MolmoAct 2-Bimanual YAM dataset: 700+ hours bimanual demos (towel folding, grocery scanning, phone charging, table bussing) curated with Cortex AI, 30x more data than original MolmoAct
- Training data mix: SO-100/SO-101 low-cost arms, DROID Franka real-world data, Google Robot BC-Z and Fractal from Open X-Embodiment, Bridge WidowX, original MolmoAct household data
- Re-annotated robot demonstrations with open VLM, increasing unique instruction labels from ~71K to ~146K across dataset mixture
- Simulation benchmarks: 20.6% success on MolmoBot household tasks (π0.5: 10.3%), 0.443 on RoboEval bimanual benchmark (π0.5: 0.405)
- Real-world zero-shot on Franka arm: 100% apple-on-plate, 93.3% knife-in-box, 93.3% red-cube-in-tape-roll, 86.7% pipette-in-tray, averaging 87.1% versus π0.5's 45.2%
- Post-training evaluation: 97.2% average on LIBERO skill retention benchmark (98.1% for MolmoAct 2-Think), improving ~10-11 points over original MolmoAct
- Third-party Cortex AI evaluation across 5 robotics policies: MolmoAct 2 scored 0.51 average (OpenVLA-OFT 0.36, π0.5 0.32, Cosmos Policy 0.16, X-VLA 0.05), ranking first on 7 of 8 bimanual tasks
- Stanford Medicine Cong Lab piloting MolmoAct 2 in wetlab automation for CRISPR gene-editing workflows, handling sample movement and benchtop equipment operation
- Supports natural-language instructions and visual traces showing desired robot path for easier deployment without full model retraining
- Reference hardware setup: two YAM arms with overhead Intel RealSense D435 + two D405 close-up cameras for tabletop manipulation (YAM arms donated by I2RT Robotics for evaluation only)
Decoder
- Action Reasoning Model (ARM): Model class introduced by AI2 that reasons about 3D environment structure before executing physical actions, versus direct action prediction from visual input
- VLA (Vision-Language-Action model): Architecture that processes visual input and language instructions to output robot control actions in a single unified system
- Flow matching: Generative modeling technique for continuous robot actions that learns to transform noise distributions into precise action sequences
- KV-cache bridge: Architecture component that shares key-value cache between vision-language model and action expert for efficient action generation without full recomputation
- FAST tokenizer: Physical Intelligence's proprietary method for discretizing continuous robot actions; AI2 released an open reimplementation trained on their own data
- Depth perception tokens: Learned representations that encode 3D depth information from 2D images, enabling spatial reasoning about object positions and manipulation
- Open X-Embodiment: Community-aggregated robotics dataset pooling demonstrations from dozens of labs across different robot platforms and task types
- LIBERO: Benchmark measuring continual learning in robotics—how well models acquire and retain multiple skills over time without catastrophic forgetting
Original article
MolmoAct 2 is an upgraded action reasoning model that improves real-world robot task performance and is paired with a large open bimanual manipulation dataset.
Explore all 66 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi