UW PhD Joins NUS, CoRL 2026 Spotlight: $6K Robot Achieves Embodied Reasoning

Career Move and New Roles

In August 2025, Dr. Jianfei Duan (段佳飞) joined the National University of Singapore (NUS) School of Computing as a Presidential Young Professor. He established the MAGIC Lab and serves as a Principal Investigator at the Singapore Embodied AI Centre (EAIC).

Education and Early Career

Duan earned his B.Eng. in Electrical and Electronic Engineering from Nanyang Technological University (NTU) with Highest Distinction. He then pursued a Ph.D. in Computer Science and Engineering at the University of Washington, co-advised by Dieter Fox and Ranjay Krishna, graduating in 2026. During his doctorate, he researched at the Allen Institute for AI (AI2) and interned as a research scientist at NVIDIA.

Research Focus

His work centers on robot learning, embodied AI, and robot foundation models, aiming to give robots perception, reasoning, and action capabilities. A core thread is converting visual information into physical actions.

Key Publications and Contributions

Foundational Survey

Before his Ph.D., he co-authored "A Survey of Embodied AI: From Simulators to Research Tasks" , which has accumulated over 800 citations. The survey organizes simulators, research tasks, and evaluation methods in embodied AI.

Task Execution and Failure Diagnosis

Manipulate-Anything (CoRL 2024) : Explores using vision-language models to drive real robots in task execution.

AHA (ICLR 2025) : Investigates how to detect robot manipulation failures and analyze their causes.

SAM2Act (ICML 2025) : Combines visual foundation models with memory architectures to assist robot manipulation.

These works address distinct challenges: task execution, failure diagnosis, and spatial information processing.

Generalization Benchmarks and Simulation

The Colosseum (RSS 2024 Oral) : Provides a benchmark for evaluating generalization of robot manipulation across diverse objects and environments.

AR2-D2 (CoRL 2023) : Explores training robot policies in simulation.

Vision-Language-Action Integration and Low-Cost Hardware

In 2025, the MolmoAct project began integrating vision, language, and action for task understanding and execution. In 2026, MolmoAct2 (CoRL 2026 Spotlight) advanced this toward complex embodied reasoning and real-world deployment. Notably, the primary training data was collected on a robot assembled from off-the-shelf parts costing under $6,000.

Open Environments and Zero-Shot Simulation

MolmoSpaces (RSS 2026 Oral) : Offers open environments and evaluation infrastructure for robot navigation and manipulation.

MolmoB0T : Explores large-scale simulation for zero-shot robot manipulation.

Awards and Recognition

Duan's papers have been selected as Spotlight or Oral at ICLR, ICML, RSS, CoRL, ICRA, and NeurIPS, winning several best-paper awards. MolmoB0T was a CoRL 2026 Oral and received multiple best-paper awards at ICRA 2026 workshops. He was also a finalist for the CVPR 2026 Spatial Intelligence Rising Stars, with coverage from MIT Technology Review and GeekWire.

Future Direction

From his UW doctorate to founding MAGIC Lab at NUS, Duan is extending his focus to real-world deployment of robot foundation models.