Sergey Levine
As seen in:
Medium,
arXiv,
IEEE Spectrum,
Substack,
DEV Community,
Robohub,
Crypto Briefing,
Tiger Sniff,
Techin World Live
Is this you? As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.
Claim your profile
Articles by Sergey Levine
Stop opening a second app to rewrite one sentence on Mac
Most AI writing tools still make you do the same annoying loop: Select text in the app where you are writing Copy it Open a chat tab or separate writing app Explain what you want Copy the result back Fix the formatting that got weird Try to remember where you were That is fine for a full essay. It is terrible for the small edits that happen 50 times a day.
A private Grammarly alternative for Mac should work inside the text field
Most writing tools make the same tradeoff. They are useful, but they pull your words away from the place you are already writing. You write in Mail, Notes, Slack, Linear, a browser textarea, or some internal company tool. Then the workflow becomes: Select the rough sentence Copy it Open another tab or app Paste it into an AI chat Explain what you want Copy the result back Check whether the tone still sounds like you That works, but it is not a great Mac writing assistant workflow.
I wanted AI rewriting to feel like a Mac keyboard shortcut
If you use AI to clean up writing on a Mac, the annoying part is usually not the model. It is the workflow. You write something in Mail, Notes, Linear, Slack, a browser textarea, or an editor. Then you copy it into ChatGPT, explain what you want, wait, copy the result back, and check whether formatting survived. That is fine for a long document.
Sergey Levine: General robotic foundation models may outperform narrow solutions, the future of medicine involves autonomous robots, and the importance of understanding physical interactions | Invest Like the Best
Key takeaways Developing general robotic foundation models may be more efficient than creating narrow, domain-specific solutions. Broader data sources can enhance the performance of robotic foundation models compared to narrow solutions. Understanding the world is crucial for robotics to generalize across tasks and systems. A general-purpose embodied foundation model could trigger a rapid expansion of robotics applications.
Physical Intelligence公司专注机器人操作系统,估值达385亿元
最近,美国具身智能公司Physical Intelligence的联合创始人Sergey Levine参加了播客节目《Invest Like The Best》,在1个多小时的访谈中,Sergey Levine就机器人模型路线、数据飞轮、本体硬件等问题进行了详细探讨。 作为一家主做软件的具身公司,PI的商业模式是给其他机器人团队提供模型,这点也类似于给汽车提供智驾系统的软件公司。并且在去年底完成6亿美元融资后,PI的估值达到56亿美元(约合人民币385亿元),比不少全栈自研的具身公司估值都要高。 另外,身为PI的联创,Sergey Levine同时也是UC伯克利大学计算机副教授,其谷歌学术引用量达18万次,在机器人等领域具有一定的影响力。 在行业共同关注的数据领域,Sergey Levine认为,现阶段量化终极数据集的规模不是首要问题,打造能落地执行多样化任务、持续采集数据的实用系统是更关键的事情。 而对于机器人大规模进入家庭这件事,Sergey...
Computer Science > Machine Learning
[Submitted on 1 Apr 2026 ( v1 ), last revised 28 Oct 2026 (this version, v2)] Title:Bridging the Simulation-to-Experiment Gap with Generative Models using Adversarial Distribution Alignment View a PDF of the paper titled Bridging the Simulation-to-Experiment Gap with Generative Models using Adversarial Distribution Alignment, by Kai Nelson and 3 other authors View PDF HTML (experimental) Abstract:A fundamental challenge in science and engineering is the simulation-to-experiment gap.
[2512.08333] Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Natural Language Actor-Critic: Scalable Off-Policy Learning in Language Space
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
[2511.07730] Multistep Quasimetric Learning for Scalable Goal-conditioned Reinforcement Learning
arXiv:2511.07730 (cs) View PDF HTML (experimental) Subjects: Machine Learning (cs.LG); Robotics (cs.RO) Cite as: arXiv:2511.07730 [cs.LG] (or arXiv:2511.07730v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2511.07730 Submission history From: Vivek Myers [ view email] [v1] Tue, 11 Nov 2025 01:28:10 UTC (8,060 KB) Bibliographic Tools Bibliographic Explorer Toggle Bibliographic Explorer () Connected Papers Toggle Connected Papers (What is Connected Papers?) Litmaps Toggle Litmaps...
[2511.00222] Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Behavioral Exploration: Learning to Explore via In-Context Adaptation
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Language Models in Plato's Cave
From its original inception, the study of artificial intelligence has been intertwined with the quest to understand human intelligence.
[2506.01716] Self-Challenging Language Model Agents
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
[2410.18076] Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration
Open access is only possible with YOUR support. Give to arXiv this week to help keep science open for all.
Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion and Aviation
Coming this September, the Forum is free, virtual, and open to all. Sign Up and Learn more.
The Global Project to Make a General Robotic Brain Original
The generative AI revolution embodied in tools like ChatGPT, Midjourney, and many others is at its core based on a simple formula: Take a very large neural network, train it on a huge dataset scraped from the Web, and then use it to fulfill a broad range of user requests. Large language models (LLMs) can answer questions, write code, and spout poetry, while image-generating systems can create convincing cave paintings or contemporary art.
[2311.05584] Zero-Shot Goal-Directed Dialogue via RL on Imagined Conversations
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Grow Your Limits: Continuous Improvement with Real-World RL for Robotic Locomotion
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Pre-training generalist brokers utilizing offline reinforcement studying – Google AI Weblog - Techin World Live
Reinforcement studying (RL) algorithms can be taught expertise to unravel decision-making duties like taking part in video games, enabling robots to select up objects, and even optimizing microchip designs. Nonetheless, working RL algorithms in the actual world requires costly energetic information assortment.
Offline RL and Large Language Models
Perhaps one of the most impressive demonstrations of the potential for large language models is their ability to hold a reasonable conversation with humans, providing responses that follow the flow of the conversation, respond to questions, and even perform rudimentary tasks, such as searching the web.
General-Purpose Pre-Trained Models in Robotics
The impressive generalization capabilities of large neural network models hinge on the ability to integrate enormous quantities of training data. This presents a major challenge for most downstream tasks where data is scarce. As a result, we have seen a transformation over the years away from training large models entirely from scratch, and toward methods that utilize finetuning or few-shot learning.
Self-Improving Robots and the Importance of Data
There is a popular idea that, with 10,000 hours of experience, a person could achieve mastery at whatever task they set themselves to. A bit of back-of-the-envelope calculation tells us that so far, our machine learning models are quite a bit more data-hungry. For example, GPT-3 is trained on over 10 terabytes of data, which corresponds to some billions of pages of text that, at average human reading speed, would take hundreds of millions of hours to read (about a hundred lifetimes).
Making RL tractable by learning more informative reward functions: example-based control, meta-learning, and normalized maximum likelihood - ΑΙhub
Diagram of MURAL, our method for learning uncertainty-aware rewards for RL. After the user provides a few examples of desired outcomes, MURAL automatically infers a reward function that takes into account these examples and the agent’s uncertainty for each state. By Abhishek Gupta, Kevin Li, Sergey Levine Although reinforcement learning has shown success in domains such as robotics, chip placement and playing video games, it is usually intractable in its most general form.
Understanding the World Through Action: RL as a Foundation for Scalable Self-Supervised Learning
Machine learning systems have mastered a breadth of challenging problems in domains ranging from computer vision to speech recognition and natural language processing, and yet the question of how to design learning-enable systems that match the flexibility and generality of human reasoning remains out of reach.
What can I do here? Learning new skills by imagining visual affordances - AIhub.org connecting the AI community and the world. - Association for the Understanding of Artificial Intelligence
By Alexander Khazatsky, Ashvin Nair, and Sergey Levine How do humans become so skillful? Well, initially we are not, but from infancy, we discover and practice increasingly complex skills through self-supervised play. But this play is not random – the child development literature suggests that infants use their prior experience to conduct directed exploration of affordances like movability, suckability, graspability, and digestibility through interaction and sensory feedback.
What Can I Do Here? Learning New Skills by Imagining Visual Affordances
How do humans become so skillful? Well, initially we are not, but from infancy, we discover and practice increasingly complex skills through self-supervised play. But this play is not random - the child development literature suggests that infants use their prior experience to conduct directed exploration of affordances like movability, suckability, graspability, and digestibility through interaction and sensory feedback.
[2106.00671] What Can I Do Here? Learning New Skills by Imagining Visual Affordances
[Submitted on 1 Jun 2021 (v1), last revised 13 Jun 2021 (this version, v2)] Authors:Alexander Khazatsky, Ashvin Nair, Daniel Jing, Sergey Levine Download PDF Abstract: A generalist robot equipped with learned skills must be able to perform many tasks in many different environments. However, zero-shot generalization to new settings is not always possible. When the robot encounters a new environment or object, it may need to finetune some of its previously learned skills to accommodate this change.
[2006.04779] Conservative Q-Learning for Offline Reinforcement Learning
[Submitted on 8 Jun 2020 (v1), last revised 19 Aug 2020 (this version, v3)] Authors:Aviral Kumar, Aurick Zhou, George Tucker, Sergey Levine Download PDF Abstract: Effectively leveraging large, previously collected datasets in reinforcement learning (RL) is a key challenge for large-scale real-world applications. Offline RL algorithms promise to learn effective policies from previously-collected, static datasets without further interaction.
RoboNet: Large-Scale Multi-Robot Learning
10/24/2019 ∙ by Sudeep Dasari, et al. ∙ 7 ∙ share Robot learning has emerged as a promising tool for taming the complexity and diversity of the real world. Methods based on high-capacity models, such as deep networks, hold the promise of providing effective generalization to a wide range of open-world environments. However, these same methods typically require large amounts of diverse training data to generalize effectively.
Visual model-based reinforcement learning as a path towards generalist robots
By Chelsea Finn∗, Frederik Ebert∗, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine With very little explicit supervision and feedback, humans are able to learn a wide range of motor skills by simply interacting with and observing the world through their senses.
Dexterous Manipulation with Deep Reinforcement Learning: Efficient, General, and Low-Cost
Abstract: Dexterous multi-fingered robotic hands can perform a wide range of manipulation skills, making them an appealing component for general-purpose robotic manipulators. However, such hands pose a major challenge for autonomous control, due to the high dimensionality of their configuration space and complex intermittent contact interactions.
Deep reinforcement and meta-learning: Building flexible and adaptable machine intelligence - Artificial Intelligence Conference in San Francisco 2018
Deep learning methods have achieved impressive results across a range of passive perception domains, from computer vision to speech recognition and natural language processing. However, these successes have been limited to problems that satisfy two critical properties: first, they rely on the availability of large amounts of labeled data, and second, they rely on the assumption that test points are independent and uniform samples rather than time steps in a sequential decision process.
[1805.12114] Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models
Abstract: Model-based reinforcement learning (RL) algorithms can attain excellent sample efficiency, but often lag behind the best model-free algorithms in terms of asymptotic performance, especially those with high-capacity parametric function approximators, such as deep networks. In this paper, we study how to bridge this gap, by employing uncertainty-aware dynamics models.
DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills
(Submitted on 8 Apr 2018 (v1), last revised 27 Jul 2018 (this version, v3)) Submission history From: Xue Bin Peng [view email] [v1] Sun, 8 Apr 2018 17:04:58 GMT (8962kb,D) [v2] Sat, 16 Jun 2018 20:48:52 GMT (8711kb,D) [v3] Fri, 27 Jul 2018 03:44:10 GMT (17423kb,D)
[1502.05477] Trust Region Policy Optimization
arXiv:1502.05477 (cs) [Submitted on 19 Feb 2015 (v1), last revised 20 Apr 2017 (this version, v5)] View PDF Comments: Subjects: Machine Learning (cs.LG) Cite as: arXiv:1502.05477 [cs.LG] (or arXiv:1502.05477v5 [cs.LG] for this version) https://doi.org/10.48550/arXiv.1502.05477 Submission history From: John Schulman [ view email] [v1] Thu, 19 Feb 2015 06:44:25 UTC (547 KB) [v2] Mon, 18 May 2015 14:56:50 UTC (540 KB) [v3] Mon, 8 Jun 2015 10:47:03 UTC (540 KB) [v4] Mon, 6 Jun 2016 01:00:57 UTC...
Show More
loading
Actions
Is this you?
As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.Get in touch with Sergey
Contact Sergey, search articles and posts on X, monitor coverage, and track replies from one place.
Learn more about Muck Rack