Insider Brief
- Allen Institute for AI, or Ai2, released a new open-source robotics model called MolmoAct 2 along with what it describes as the largest open-source dataset for two-armed tabletop robot manipulation.
- Ai2 said MolmoAct 2 is designed to help robots perform real-world manipulation tasks such as moving objects, handling laboratory tools and folding towels while operating faster and across more tasks than its earlier system without extensive retraining.
- The project included outside contributions from Cortex AI for independent benchmarking, I2RT Robotics for robotic hardware used in testing and Stanford School of Medicine for evaluating the system in CRISPR laboratory workflows tied to automated wetlab operations.
AI research group Allen Institute for AI, or Ai2, announced it is releasing a new open-source robotics model and large training dataset aimed at improving how robots perform physical tasks in real-world environments.
The updated system, called MolmoAct 2, is designed to help robots better understand spatial environments and respond to instructions while performing manipulation tasks such as moving objects, folding towels, handling laboratory tools and clearing tables. According to the non-profit organization, the model operates significantly faster than its earlier version and is intended to work across multiple tasks without extensive retraining.
Alongside the model release, the group published what it describes as the largest open-source dataset focused on two-armed tabletop robot manipulation. The dataset includes more than 700 hours of demonstrations involving coordinated robotic tasks such as scanning groceries, charging smartphones and organizing objects.
The non-profit said one goal of the project is to make robotics research more transparent at a time when many advanced robotics AI systems remain closed or partially inaccessible to outside researchers. The release includes model weights, datasets and supporting tools intended for academic and commercial experimentation.
MolmoAct 2 combines visual reasoning with robotic action planning, allowing systems to process images, spatial relationships and instructions before generating movement commands. The organization said the model can also selectively apply deeper 3D reasoning in situations where additional spatial understanding is needed while avoiding unnecessary computational overhead.
According to internal testing cited by Ai2, the system performed well across both simulated and real-world robotics benchmarks involving object placement, manipulation and coordinated bimanual tasks.
Ai2’s development of MolmoAct 2 included:
- Cortex AI conducted an independent benchmark evaluation of MolmoAct 2 across multiple real-world bimanual robotics tasks. The company compared MolmoAct 2 against several competing robotics models using repeated task trials and systematic scoring methods.
- I2RT Robotics provided YAM robotic arms that were used during some of the in-lab evaluations of MolmoAct 2. According to the company, I2RT did not participate in developing the model, designing the tests or reporting the results.
- Stanford School of Medicine partnered with the project to test MolmoAct 2 in laboratory workflows tied to CRISPR gene-editing experiments. Researchers used the system for routine lab manipulation tasks such as moving samples and operating equipment in an effort to explore more automated “self-driving” wetlab operations.
Ai2 acknowledged that the system still faces limitations common in robotics, including challenges with fine-grained manipulation, camera occlusion and response timing during complex physical tasks.
“These are exactly the kinds of challenges that shared foundations can help the field tackle — models researchers can inspect, datasets they can build on, and (coming soon) training code they can adapt to new machines and situations,” the organization noted in the announcement. “MolmoAct 2 is meant to help set that standard, building on the groundwork we laid last year with MolmoAct and translating our earlier research into tangible impact.”
Allen Institute for AI, commonly known as Ai2, is a nonprofit artificial intelligence research institute founded in 2014 by Microsoft co-founder Paul Allen. The organization conducts AI research and engineering focused on open science and public-interest applications, with operations based in Seattle and an additional office in Tel Aviv.
