Asian Giant Bets on Embodied AI

The National University of Singapore (NUS) has long ranked among Asia's top universities and is now strengthening its position in the field of embodied AI — artificial intelligence that physically interacts with the world. The new MAGIC Lab (Manipulation and General Intelligence Control Lab) is being established under the umbrella of NUS Computing, a school with roots dating back to 1975, when it began teaching computer science as part of Nanyang University. The NUS School of Computing was officially founded in 1998 and today is one of the world's leading institutions in the field.

The new lab is led by Jiafei Duan, who joins as a Presidential Young Professor. Duan is currently completing his PhD at the University of Washington in Seattle under the guidance of renowned professors Dieter Fox and Ranjay Krishna. His career includes experience at NVIDIA, Singapore's A*STAR research institute, and the non-profit Allen Institute for AI (Ai2), where he led the development of the MolmoAct model.

MolmoAct: When AI Starts to "Think" in 3D

The MolmoAct-7B model, created by Duan and his team at Ai2, represents a new category of so-called Action Reasoning Models (ARM). Unlike classical language models that reason in text, MolmoAct can spatially reason about actions. For a layperson, one can imagine that instead of describing "pick up the bottle," the model actually plans the trajectory of movement in three-dimensional space, estimates distances between objects, and translates this into motor commands for a robotic arm.

According to the Ai2 technical report, the model works in three steps: first, it creates spatial perception of the scene using special tokens with depth information, then it plans movement as a sequence of points in the image, and finally decodes actions for specific hardware — whether it's a robotic arm or a humanoid robot.

The results are impressive. On the SimplerEnv benchmark, which tests robots' ability to handle tasks outside training data, MolmoAct achieved a success rate of 72.1%, surpassing models from Physical Intelligence, Google, Microsoft, and NVIDIA. In the LIBERO simulator, focused on transferring knowledge between different tasks, the model achieved an average success rate of 86.6%. The training efficiency is also noteworthy: pre-training on 256 NVIDIA H100 GPUs took approximately one day, while fine-tuning on 64 H100s took just two hours.

MAGIC Lab and Its Mission

The newly established lab has an ambitious goal: to develop the "next generation of human-centric models for robotic manipulation" — that is, models designed to be safe, reliable, and easily deployable in the real world. Its focus includes several key areas:

• Multimodal language models for robotic reasoning
• 3D vision and spatial understanding
• Robotic learning from data and simulations
• Simulation as a bridge between the virtual and physical world
• Dexterous manipulation
• Cross-embodiment learning — the ability to transfer knowledge between different types of robots

The lab is currently seeking PhD students, research assistants, and postdocs in the fields of robotics, computer vision, machine learning, and related disciplines. Thus, NUS Computing joins the global race for talent in embodied AI, alongside institutions such as Stanford, MIT, Carnegie Mellon, and ETH Zurich.

Why This Matters for Europe and the Czech Republic Too

The development of embodied AI is not merely an academic matter. According to analysts' estimates, the autonomous robotics market will more than double by 2033, and artificial intelligence capable of physical interaction will be key for manufacturing, logistics, and healthcare. Singapore, which has already invested in humanoid robots and satellites, confirms its strategy to become a global hub for AI and robotics with this step.

For the European scene — and thus the Czech Republic — it is crucial that Duan's MolmoAct is a fully open model. Not only the model itself was published, but also training data, code, evaluation scripts, and checkpoints. This is a key advantage at a time when commercial giants like OpenAI or Google often keep their models closed. Czech universities and research institutions, such as groups at CTU or Masaryk University, can use open models like MolmoAct as a foundation for their own research without licensing fees.

At the same time, the regulatory context must be mentioned: the EU AI Act, which is coming into full force, classifies robotics with autonomous decision-making as a high-risk area. Research into transparent, explainable models — such as MolmoAct presents itself through the visualization of its "thought processes" in the image — can help European companies meet strict requirements for the safety and auditability of AI systems.

Openness as Asia's Competitive Advantage

While the United States bets on commercial closed systems and China massively invests in state-backed humanoid robots, Singapore has chosen the path of open research. By combining the world-class academic environment of NUS with experience from American institutions (UW, Ai2) and industry (NVIDIA), MAGIC Lab creates an ecosystem that can attract talent from around the world — including Europe.

Jiafei Duan emphasized in his announcement that the goal is to build models "designed to be deployed safely, reliably, and easily in the real world." This is exactly the approach the European market needs: not another chatbot, but AI that actually does something physically — while being transparent enough for regulators and users to trust it.