Ai2 Open-Sources MolmoAct 2 Robot Foundation Model, Outperforms Physical Intelligence π0.5
The Allen Institute for AI (Ai2) fully open-sourced MolmoAct 2 on May 5, 2026 — releasing model weights, training code, and complete training data simultaneously. The 7B-parameter robot foundation model outperforms Physical Intelligence's closed-source π0.5 across simulation and real-world benchmarks.
Performance
On real-world DROID tasks in zero-shot settings, MolmoAct 2 achieves up to 87.1% success. Task-level numbers:
- Apple-to-plate transfer: 100% success
- Pipette-to-tray: 86.7% success
- Multi-step object sorting: 62% success
Architecture
MolmoAct 2 couples the Molmo 2-ER vision-language model with a dedicated action expert using flow matching, connected via a KV-cache bridge. This design delivers a 37× inference speedup over prior architectures.
Open Dataset
Training uses 12,000 real-world robot episodes. Ai2 also released MolmoAct 2-Bimanual YAM, a massive open-source bimanual tabletop manipulation dataset with over 700 hours of demonstrations.
Details are in the Ai2 blog post and arXiv paper.
Related Articles
Carnegie Mellon University and the Bosch Center for AI developed HTD (Humanoid Transformer with Touch Dreaming), which trains robots to anticipate future tactile signals. Across five dexterous real-world tasks, HTD achieved 90.9% higher success rates over vision-only baselines.
NVIDIA released open-source physical AI agent skills across Omniverse, Cosmos, Isaac, Metropolis, Alpamayo, and Jetson. The company points to manufacturing gains including 67% faster training and deployment at Pegatron and a 17% detection-rate improvement at Delta Electronics.
Figure’s robot fleet has crossed its human headcount for the first time, according to CEO Brett Adcock. The chart shared with the tweet points to a steep 2026 production ramp and a workforce line near the mid-hundreds.