Beijing physical-AI startup founded 2025; its Liber-0 world action model, pretrained on 2.2M+ hours of human data, hit 36% zero-shot task success.
LiberAI is a physical AI company in Beijing that trains robot manipulation foundation models on recorded human experience instead of on large fleets of teleoperated robots. It was founded in December 2025 by Songming Liu, a Tsinghua University PhD candidate who previously led the RDT series of robot foundation models (RDT-1B in 2024 and RDT-2 in 2025), with co-founder Fanqi Lin, a Tsinghua PhD student whose earlier work includes a data scaling laws study for robot imitation learning. By October 2026 the company employed more than 130 people. Four months after founding it released its first foundation model for high-degree-of-freedom dexterous hands.
Its main model line is Liber-0, described as a world action model: a Mixture-of-Transformers network with language, vision and action experts that jointly predicts future video frames and robot actions, conditioned on a Unified Embodiment Representation that encodes the geometry and kinematics of the body carrying out the task. LiberAI says Liber-0 was pretrained on more than 2.2 million hours of human experience, mixing web-scale human video with glove recordings and robot data, and co-trained on visual question answering, segmentation and video understanding. In a zero-shot evaluation published in November 2026, covering 60 tasks proposed by GPT-6 and 600 trials, it reported 36.0% overall success: 63.5% on pick-and-place, 34.0% on other manipulation and 10.5% on long-horizon tasks, without robot demonstrations collected for those tasks. A preview version released on 17 September 2026 placed first on the RoboDojo leaderboard at launch with an average score of 30.74 across 42 tasks.
Two pieces of in-house hardware support the model. LiberSense Glove is a wearable data-collection system 0.3 mm thick and weighing 131 g that measures finger motion and synchronises it with a first-person camera and two wrist cameras, so demonstrations can be gathered without a robot on site; the company said in July 2026 that the gloves had entered mass production. LEO is a robot platform with dexterous hands, slim forearms, a wrist with three rotational degrees of freedom and an articulated waist, whose chest and wrist cameras mirror the glove-collection viewpoints. In company experiments, 10,000 hours of glove data outperformed 2 million hours of web video, and combining glove data with a quarter of the robot demonstrations beat training on the full robot set. A browser demo lets users upload an object for Liber-0 to manipulate in simulation, the company has shown the model preparing boba tea, and its stated long-term aim is general-purpose home robots.
LiberAI disclosed four financings within five months in 2026: seed, angel and angel+ rounds totalling nearly RMB 500 million backed by ZhenFund, HongShan, Meituan Long-Z and Shunwei Capital; a Pre-A round led by Shunwei Capital in June; a Pre-A+ round including Sinovation Ventures and CMC Capital in July; and a Pre-A++ round in September that took cumulative funding past RMB 1 billion.
Primary type & automation activities this supplier delivers:
World action model pretrained on 2.2M+ hours of human experience for zero-shot dexterous robot manipulation.
Lightweight data glove that records measured finger motion with first-person and wrist video for robot training.
Human-aligned robot with dexterous hands, 3-DoF wrists and an articulated waist for human-to-robot skill transfer.
Contact LiberAI
WEBSITE
https://www.liberai.comHEADQUARTERS
China
Company Facts
Founded
2025
Primary Role
Software/Algorithm
Company Size
-
Primary Region
Asia-Pacific
Annual Sales
-
Funding Stage
Early-Stage VC
Funding Total
-