May 26, 2025 — As embodied AI moves from eye-catching demonstrations toward real industrial use, Jianlan Luo is emerging as one of the researchers most closely associated with that transition. The University of California, Berkeley postdoctoral scholar and robotics specialist, who was named to the 2024 MIT Technology Review TR35 list, has built his reputation around a central question: how to make embodied intelligence not just impressive, but genuinely useful.
Luo’s work focuses on a problem that still defines much of the robotics industry. Many robots today can walk, jump, or perform carefully staged motion demos, yet relatively few can carry out demanding manipulation tasks in uncertain physical environments. In Luo’s view, the industry’s most serious shortfall is not movement itself, but manipulation — the ability to use robotic hands to recognize different materials, grasp irregular objects, and complete precise operations in open, dynamic settings.
That emphasis on manipulation has shaped more than a decade of Luo’s research at the intersection of robot control and reinforcement learning. During his doctoral studies at UC Berkeley, he worked to build learning frameworks grounded in the physical world rather than idealized environments. After returning to the Berkeley Artificial Intelligence Research Lab as a postdoctoral researcher in 2022, he led the development of HIL-SERL, a real-robot reinforcement learning system designed to deliver broad generalization across multiple tasks.
In one of the demonstrations highlighted in the article, a robot armed with a whip-like tool extracts a specific wooden block from a tightly stacked tower with smooth, stable motion. The task may look effortless on video, but it demands an extraordinary level of control and sensitivity to physical uncertainty. According to the article, the robot in the HIL-SERL system reached a 100% success rate after only two hours of training.
Luo describes manipulation as the crown jewel of robotics because it requires more than movement. A capable embodied system must perceive uncertainty, model physical interaction quickly, and respond with dexterity under real-world constraints. That is why HIL-SERL departs from the simulation-heavy path followed by many robotics projects. Instead of depending primarily on synthetic environments, the system is built around real-world robot data, forcing the learning process to engage directly with the complexity, noise, and unpredictability of physical interaction.
For Luo, that choice is fundamental rather than stylistic. Simulators, he argues, cannot fully reproduce the richness of the real world. True embodied intelligence must be shaped by direct physical feedback, where each success and failure becomes part of the learning process. In this view, manipulation is not simply an engineering subproblem; it is one of the clearest tests of whether a machine can actually understand and interact with the world around it.
The article also traces Luo’s experience across both academia and industry, including periods at UC Berkeley and Google X. He has consistently argued that research and industrial deployment should form a positive flywheel rather than separate tracks. Real-world scenarios expose weaknesses in algorithms, while algorithmic advances make larger-scale deployment possible. Luo has urged young researchers not to focus solely on publishing top papers, but to ask whether the underlying problem is truly worth solving and whether the solution can survive contact with a real system.
That perspective informs his broader view of what general-purpose humanoid robots still lack. In Luo’s assessment, the missing piece is closed-loop capability — the ability to perform a task, learn from each interaction, and immediately improve. A robot that fails during assembly, for example, should not have to return to simulation and start over. It should be able to recognize what went wrong, adjust its strategy in the moment, and carry that lesson forward. Over the next decade, Luo believes, genuinely useful embodied AI systems will need that kind of memory and self-improving feedback loop.
Luo has since joined AGIBOT as Chief Scientist, where he is helping build systems that connect embodied intelligence research with real deployment. His current work includes integrated cognitive-and-control architectures, platforms for data collection, training, and inference, and embodied interaction technologies for dexterous robotic hands. The broader goal, he has said, is to make robots operate in the real world 24 hours a day, seven days a week, no longer as laboratory demonstrations but as useful systems capable of working in factories, unmanned convenience stores, and hazardous environments.
For Luo, the significance of recognition such as TR35 lies less in the honor itself than in the responsibility that follows. The past decade, he argues, was about proving that reinforcement learning could run on real robots. The next decade will be about proving that it can create real value.
