PsiBot recently completed its angel round financing, led by GL Ventures and Lanchi Ventures.
With the new funding, PsiBot will continue advancing the training of robot skill sets based on reinforcement learning algorithms, scenario-driven data generation and collection, and the development and deployment of end-to-end solutions—aiming to build the industry’s leading general-purpose dexterous manipulation agents.
Highest Density of Product Veterans and Scientists
Dr. Viktor WANG, founder of PsiBot, brings nearly two decades of proven leadership in mobile devices, smart speakers, and robotics. He has successfully driven multiple product lifecycles from definition and development to launch and global expansion, building complete 0-to-1-to-N commercialization loops—earning him the reputation of a seasoned “veteran” in robotic productization.
Co-founder Dr. Xiaojie CHAI has over 15 years of experience in robotics and autonomous driving, with expertise spanning algorithms, simulation, engineering, and full-stack development. He has led L4 product deployment with closed-loop data integration. Dr. Chai is a seasoned R&D expert with extensive experience in scalable production.
PsiBot is also recognized as the “embodied AI company with the highest density of scientists.” The company has established the PKU-PsiBot Joint Lab for Embodied Dexterous Manipulation, led by Prof. Yaodong YANG from the Institute for Artificial Intelligence at Peking University.
Prof. Yang, a leading young scholar in reinforcement learning and a Ph.D. graduate from UCL—the birthplace of modern deep RL—has spearheaded several groundbreaking contributions. These include leading the first Chinese team to publish multi-agent RL research in Nature Machine Intelligence and winning the NeurIPS 2022 Embodied Dexterous Manipulation Challenge.
The joint lab also collaborates with Prof. Yitao LIANG on research into long-horizon planning for embodied agents. Prof. Liang has long focused on enhancing machine learning through structured knowledge integration and has achieved significant results in open-ended environments such as Minecraft. His work uses large language models (LLMs) to break down and plan complex tasks, enabling AI “brains in jars” to develop hands and feet for autonomous action in embodied environments.
Beyond the joint lab, PsiBot’s co-founder Yuanpei CHEN—an exceptionally talented robotics prodigy born after 2000—is also instrumental to the team. A visiting scholar at Stanford under the mentorship of Professors Karen Liu and Fei-Fei Li, he was the first to simultaneously control bimanual, multi-skill manipulation in the real world using reinforcement learning.
Another key technical team member is Prof. Ying WEN from the School of Artificial Intelligence at Shanghai Jiao Tong University. His group developed the multimodal decision-making model DB1, which outperformed DeepMind’s generalist Gato model. By introducing over a hundred real-world tasks, DB1 serves as a powerful tool for bridging foundational AI research with practical applications.
Led by seasoned product veterans and powered by a uniquely dense team of scientists, PsiBot has formed a “7890 hexagonal task force” that spans four generations—those born in the ’70s, ’80s, ’90s, and ’00s. The team combines deep technical know-how, product insight, and strong execution. Just like the humanoid robots they build using reinforcement learning, PsiBot is distinguished by cutting-edge innovation, vast commercial potential, and formidable capability.
The "Impossible Triangle" of Embodied Intelligence
High generalization, high dexterity, and high success rate form the “Impossible Triangle” of embodied intelligence.
High generalization refers to a robot’s ability to perform complex tasks across diverse and changing environments and objects. High dexterity reflects the precision and flexibility required for fine tasks such as assembling LEGO bricks or bimanual manipulation. High success rate ensures that the robot consistently and correctly completes tasks even under disturbances—typically requiring a 95% success rate during validation and over 99.9% for scaled deployment. Balancing all three is extremely challenging.
Achieving high generalization demands a universal model and learning algorithms that embrace data diversity. High dexterity relies on fine-tuned models, precision-focused learning, and advanced control algorithms. Robustness further raises requirements on the control system’s stability and adaptability.
Reinforcement learning (RL) is the cornerstone for enhancing all three aspects. It enables agents to train with low-cost synthetic data in simulation environments, perform autonomous exploration, and learn optimal strategies through trial and error. This not only improves dexterity—potentially beyond human capabilities—but also boosts robustness.
PsiBot has been deeply focused on full-stack RL capabilities since its inception, building core technological barriers in embodied dexterous manipulation before expanding into general-purpose manipulation.
Its hierarchical end-to-end framework leads the industry, featuring the Psi-P0 planning model and the Psi-C0 control model. Psi-P0 uses large models for reasoning and interaction to understand how actions affect the environment, allowing complex tasks to be decomposed and generalized.
Psi-C0, developed by Yuanpei Chen, adopts a dual-layer architecture that combines human motion data with deep reinforcement learning. The upper layer trains a trajectory generator using human data, while the lower layer uses these trajectories to guide RL training—solving both generalization and dexterity challenges. It also represents the world’s first RL-based system capable of controlling dual-arm, dual-hand multi-skill operations in the real world.
Psi-C0 enables sequential training for multi-skill execution. In collaboration with Stanford, Yuanpei Chen introduced Sequential Dexterity, a method for chaining multiple dexterous manipulation strategies to complete long-horizon tasks.
In real-world block assembly tasks, this approach links four distinct skills—searching, reorienting, grasping, and inserting bricks—significantly improving overall task success. This marks the world’s first long-horizon dexterous manipulation task completed via reinforcement learning, demonstrating generalization across multi-skill sequences.
Prof. Yitao Liang developed the Psi-P0 model to enable task decomposition and planning for complex missions in open environments. The complexity and accuracy of the tasks supported by Psi-P0 surpass those of contemporaneous efforts from OpenAI (VPT) and Nvidia (MineDojo).
As demonstrated in the video below, the model can be further extended by integrating memory structures, enabling embodied agents to exhibit lifelong learning—the ability to continuously improve based on their own experiences.
In terms of product strategy, PsiBot is entering the market through B2B service sectors, guided by high-value scenario demands from leading clients. The company focuses on developing and integrating skill sets to enable commercial deployment, while rapidly iterating on hardware, algorithms, and data systems. This approach continuously enhances the generalization, dexterity, and success rate of its embodied AI solutions—striving to deliver the optimal answer to the “Impossible Triangle” of embodied intelligence.
Messages from the Founder and Investors
Message from Dr. Viktor Wang, Founder and CEO of PsiBot: “We are deeply grateful for the strong support from our investors. Our team—comprising seasoned industry veterans and a high-density group of scientists—possesses a full-stack technical capability and is committed to building integrated hardware-software solutions at the forefront of embodied intelligence. We aim to expand applications across advanced manufacturing, retail logistics, and B2B service industries, rapidly closing the data loop and achieving commercial deployment. As we stand at the dawn of embodied intelligence, we look forward to growing with our partners—intelligently and purposefully—toward a smarter future.”
Message from GL Ventures: “As a deep integration of artificial intelligence and the physical world, the embodied intelligence industry holds tremendous potential and is poised to transform countless sectors. PsiBot has made significant progress in solving the ‘Impossible Triangle’ of high dexterity, generalization, and success rate, thanks to its strong foundation in reinforcement and imitation learning. The founding team brings together top talent from multiple domains, combining technical excellence with outstanding engineering and commercialization capabilities. We believe that under the leadership of Dr. Viktor WANG, PsiBot will deliver revolutionary applications and efficient solutions, driving both technological breakthroughs and business value.”
Message from Lanchi Ventures: “The embodied robotics market holds immense promise. We see general-purpose manipulation for complex tasks as a key technical bottleneck for bringing embodied intelligence into the real world. PsiBot’s team is among the world’s best in reinforcement learning and embodied AI. More importantly, they deeply understand both industry needs and application scenarios, with strong system architecture design capabilities, hands-on product deployment experience, and supply chain advantages. As the technology matures and the industrial ecosystem evolves, we expect the embodied robotics market to enter a phase of rapid growth. We are highly optimistic about PsiBot’s global potential.”