Exclusive Interview with Physical Intelligence Scientist Ke Liyiming: Inside the Robot Industry Frenzy, Reinforcement Learning Is the Key to Breaking Through
An English translation and commentary of the full text is provided within this substack
Member’s Commentary
Qian Yu’s (钱禹) commentary:
Liyiming Ke (柯丽一鸣) is a research scientist at Physical Intelligence π.
Physical Intelligence itself has become one of the most heavily funded robotics companies in history. Founded in 2024 by Karol Hausman and co-founded by Sergey Levine (one of the most cited researchers in reinforcement learning, with over 230,000 Google Scholar citations and an Associate Professor at UC Berkeley), the San Francisco-based startup raised a $400 million Series A in November 2024 at a $2 billion valuation, led by Jeff Bezos, OpenAI's Startup Fund, Thrive Capital, and Lux Capital. A $600 million Series B followed in late 2025 at a $5.6 billion valuation, led by CapitalG (Alphabet's growth fund). As of early 2026, the company is reportedly in talks to raise an additional $1 billion at an $11 billion valuation.
Ke completed her PhD at the University of Washington (2017-2024) under Siddhartha Srinivasa, director of the Personal Robotics Lab and one of the most influential figures in manipulation research. Her dissertation, Data-Driven Fine Manipulation, tackled what she calls the "last millimeter" problem: getting robots to perform tasks that require human-level dexterity, like using chopsticks. During her doctoral work, she held research internships at Meta AI (FAIR), Microsoft Research, and Google Search. She was also selected as an EECS Rising Star, a recognition given to early-career researchers identified as future leaders in electrical engineering and computer science.
At Physical Intelligence, Ke became a core contributor to the company's two flagship models. π0 (pi-zero), announced in October 2024, is a Vision-Language-Action (VLA) flow model that uses diffusion/flow-matching to generate continuous robot actions. It was designed as a generalist policy: a single model capable of controlling multiple robot embodiments across diverse tasks. π0.5, released in 2025, extended this with co-training on heterogeneous data sources to achieve open-world generalization, enabling robots to perform real-world tasks outside the laboratory. Both models were open-sourced through the OpenPI repository on GitHub. Ke's most recent contribution, RL Token (RLT), addresses a specific and critical gap: how to take a generalist VLA model and rapidly specialize it for precise manipulation using online reinforcement learning, requiring only hours of real-world data rather than weeks or months. More specifically, RLT extracts a compact representation from Vision-Language-Action (VLA) models which enables a small reinforcement learning module to rapidly fine-tune precise manipulation skills using only hours of real-world training data.
Ke's central argument, that reinforcement learning remains indispensable despite the current vogue for imitation learning and foundation model scaling, carries real weight. The dominant paradigm in robot learning since 2023 has been to collect massive demonstration datasets and train VLA models through behavioral cloning. Ke does not dismiss this approach but insists it is insufficient. Her RL Token work shows that a small RL module, bootstrapped from the VLA's learned representation, can close the gap between "roughly correct" and "precisely right." What Ke describes as the need for evaluation data (data generated during robot testing) to vastly exceed demonstration data echoes a principle well established in reinforcement learning: the agent must explore far beyond what it has been shown.
Exclusive Interview with Physical Intelligence Scientist Ke Liyiming: Inside the Robot Industry Frenzy, Reinforcement Learning Is the Key to Breaking Through (专访 Physical Intelligence 科学家柯丽一鸣：机器人行业狂欢，强化学习是破局关键)
Source: Weixin | Authors: Xu Huazhe (许华哲) | Date: April 2026
Author Background: The interviewer, Huazhe Xu (许华哲), is a tenure-track assistant professor at Tsinghua University’s Institute for Interdisciplinary Information Sciences (IIIS), the institution founded by Turing Award laureate Andrew Yao (姚期智). For more information and about this special institute check out this article.
Xu completed his PhD at the University of California, Berkeley (2016–2021), followed by a postdoctoral fellowship at Stanford University (2021–2022), before returning to China. His research focuses on embodied AI, reinforcement learning, and tactile sensing for robotics. He is a prominent figure in China’s AI community and is recognized as one of the “Four Berkeley Returnees” (伯克利四子). Xu is a co-founder of Galaxea AI (星海图), and more recently founded Broken Shell Intelligent Technology ((破壳智能科技), which received direct investment backing from Galaxea AI. He also serves as the head of the Tsinghua Embodied AI Lab (清华大学具身智能实验室，TEA Lab).
Xu is also currently teaching a course on Deep Reinforcement Learning at Tsinghua University for Spring 2026.
Brief outline
The robotics industry is caught between capital-fueled hype and genuine technical breakthroughs, with too many companies prioritizing PR over substance.
Ke Liyiming argues that robot commercialization must solve three core problems in parallel: research breakthroughs, hardware deployment, and productization.
Reinforcement learning remains essential for robotics, and the current methods, while still rough, are advancing rapidly through techniques like RL Token.
Active perception and fine manipulation remain the biggest technical bottlenecks limiting robot deployment in real-world environments.
Ke's personal motivation is deeply humanistic: using robotics to give introverted and independent people the freedom to live without depending on others.
Note: Bold-italicized text indicates emphasis by the translators. Photo of Ke Liyiming, research scientist at Physical Intelligence.
The robotics industry today is caught between two forces: capital chasing deals and concepts running wild on one side, explosive technical progress on the other. How to bring these two sides into alignment.
Today we are interviewing Ke Liyiming (柯丽一鸣), a scientist deeply engaged in the field of Physical Intelligence. She is a full-stack roboticist and a pioneer in reinforcement learning, with representative work including the recent RL Token and her earlier hand-built chopsticks robot. After years in the field, she has held fast to the original spirit of scientific research. This time, she speaks bluntly, puncturing certain illusions and offering a genuine direction for the robotics field to break through.
Core Viewpoint: Research Is the Foundation, Do Not Let Capital's Carnival Drown Out Technical Truth
From the perspective of a researcher in the field of Physical Intelligence, Ke Liyiming has no interest in "traffic-driving formulas." She deals only in solid technology and real-world deployment.
She has consistently maintained that breakthroughs in robotics have never come from hype and capital stacking, but require day after day of accumulated technical progress. "Many companies set their targets too aggressively while ignoring the most fundamental problems: scientific research, hardware, and productization. None of these is simple."
In her view, reinforcement learning is one of the key tools for unlocking the bottleneck of robot autonomy. "Still questioning whether reinforcement learning is useful for robotics in 2026? The answer has been clear for a long time. It is just that the current methods are not yet mature enough."
In an industry awash in buzzwords, her position is clear: the balance between research and commercialization is the core of sustainable industry development.
Exclusive Excerpts: A Scientist Willing to Speak the Truth
The following are Ke Liyiming's own statements:
"My background is in fine manipulation problems, so I am especially sensitive to 'observability.' Our current robots have terrible perception. Many tasks that humans can see clearly and perform easily, robots simply cannot do. The precision is frustratingly poor. Looking at the history, active perception may be a breakthrough point."
"Robot commercialization must solve three core problems: research breakthroughs, hardware deployment, and productization. Hardware determines what tasks can be done and how long the robot can operate without breaking. Research determines how well those tasks can be performed. Productization requires considering setup costs and human interaction. None of these three problems is easy."
"I support open source, but I do not advocate open-sourcing everything blindly. Open-sourcing everything is actually a waste of effort. Open source is sharing, but it is also a hot potato: time-consuming, labor-intensive, and with uncertain returns, especially when there is so much noise in the industry right now."
"Reinforcement learning is absolutely useful for robotics. There is no need to doubt this! It is just that the current methods are still relatively rough, and the volume of evaluation data (data generated during robot assessment) must far exceed demonstration data to truly improve model capability."
"Doing robotics right now is almost absurd. Every joint is immature and requires R&D, and R&D requires top talent. In a stable industry, each person working on each direction could start their own company. A lot of things we would prefer to outsource to upstream and downstream partners, but right now every link in the robotics chain has gaps. We just have to push through it."
"My original motivation for doing robotics is simple: I want to use technology to give people who like being alone and are introverted a kind of freedom from depending on others, the ability to live well without being forced to rely on anyone."
Keypoint Breakdown: Four Points to Understand the Truth of the Robotics Industry
Drawing on Ke Liyiming's remarks, we have organized four core takeaways:
Keypoint One: Technical Bottlenecks: Observation and Fine Manipulation Remain the Biggest Weaknesses
As a scientist who came up through fine manipulation research, Ke Liyiming cares most about the "fundamentals": perception capability and fine manipulation. She states bluntly that current robot observation systems are deeply incomplete. Many tasks humans can accomplish effortlessly leave robots completely helpless. In generalized scenarios, insufficient precision severely constrains robot deployment in homes, laboratories, and other settings.
Keypoint Two: Some Companies' PR Exceeds Their Technology, and Commercialization Is Too Aggressive
When discussing the current robotics industry, Ke Liyiming does not shy away from naming the disorder. Some companies blindly pursue commercialization hype while saying nothing about the complexity of research, hardware, and productization, especially problems of robustness, safety, and interaction that are difficult to solve in the short term. There are also individual companies that generate plenty of buzz and hot topics, but whose actual technical capabilities cannot be independently verified, and industry transparency remains low.
Keypoint Three: Reinforcement Learning Works, but Open Source Requires Rationality
As a firm supporter of reinforcement learning, Ke Liyiming has always believed that RL can enable robots to achieve self-improvement, gradually freeing humans from tedious operations.
But she also stresses that RL deployment depends on good priors and the accumulation of large volumes of evaluation data. Only when the quantity of evaluation data far exceeds demonstration data can the overall capability of the model truly improve.
On open source, her attitude is similarly measured. She recognizes its value but opposes a one-size-fits-all approach to open-sourcing. In the current industry environment, prioritizing core technology development matters more than chasing the attention that open source brings.
Keypoint Four: Using Technology to Empower Every Individual
Unlike many researchers who pursue "scaled deployment" and "commercial profitability," Ke Liyiming's original motivation for robotics is warm and simple.
She hopes that through robotics technology, people who are introverted and enjoy solitude will not need to depend on others or on groups. They can have sufficient life security and gain "freedom from having to rely on the collective." From small things like handling daily chores and custom manufacturing, to larger ambitions of using robot swarms to create a "cyberpunk pastoral idyll." Technological development can make people more free.
On entrepreneurship, she remains cautious. Every link in the current robotics field has gaps. Starting a company requires enormous resources and top talent. Rushing in blindly makes success very difficult.
Final Note: Moving Forward Rationally Is the Path to Breakthrough
The robotics industry today does not lack capital, concepts, or hype. What it lacks are people willing to settle down and do the technical work, people willing to tell the truth.
Perhaps the real breakthroughs in robotics will never come from mythology. They will come from accumulating data point by data point, tuning algorithms, and grinding on systems.
FOOTNOTES:
[1]: Active perception refers to the paradigm where a robot actively controls its sensors (e.g., adjusting camera angle or repositioning) to gather more informative observations, rather than passively receiving whatever sensory input happens to be available.