LiberAI

LiberAI is a Beijing physical AI company that trains robot foundation models on recorded human experience rather than on large fleets of teleoperated robots. Songming Liu, a Tsinghua University PhD candidate, founded it in December 2025 with co-founder Fanqi Lin. By October 2026 it employed more than 130 people. Alongside the models, the company builds its own data-collection gloves and a robot platform designed to resemble the humans it learns from.

Founders

Liu studied under Professor Jun Zhu at Tsinghua and moved from AI for physics into embodied AI during his doctorate. Before LiberAI he led the RDT series of robot foundation models: RDT-1B, released in October 2024, which paired large-scale pretraining with a diffusion transformer, and RDT-2 in September 2025, pretrained on human data gathered with UMI handheld grippers rather than robots. Lin is a PhD student at Tsinghua's Institute for Interdisciplinary Information Sciences, advised by Professor Yang Gao. His work includes the Data Scaling Laws paper on robot imitation learning, an ICLR 2025 oral, and a stint with Toyota Research Institute's Large Behavior Models team.

The bet: dexterous hands and human data

Most of the industry starts with one-degree-of-freedom grippers, because they produce working demos quickly. LiberAI went straight to multi-fingered hands: the hand it deploys has 20 degrees of freedom. Liu's argument is that human video and human hand recordings line up naturally with a five-fingered hand and never with a gripper, so a model trained this way should transfer down to grippers far more easily than a gripper model scales up to hands.

That choice also decides how the company collects data. Teleoperation, UMI grippers and exoskeletons ask a person to move like a robot. LiberAI's gloves do the opposite and try to stay unnoticed, so people move as they normally would. The gloves record hand poses and trajectories during everyday manipulation. Paired with first-person and wrist-mounted cameras, those recordings give the model action labels that ordinary video lacks. LiberAI said in July 2026 that the gloves had entered mass production. They do not yet capture touch or force, which the company lists as future work alongside collection in a wider range of real homes.

Liber-0

Four months after it was founded, LiberAI released its first foundation model for high-degree-of-freedom dexterous manipulation. Its main model line is Liber-0, a world action model that predicts future video frames together with the actions that produce them. The company says Liber-0 was pretrained on more than 2.2 million hours of human experience, mixing broad human video with glove and robot recordings. On 17 September 2026 it released Liber-0 Preview, which placed first on the RoboDojo leaderboard at launch: an average score of 30.74 and an average success rate of 25.52% across 42 tasks. That was the day after GPT-6 Astra posted its robotics results, and LiberAI's announcement placed Astra third.

Funding

LiberAI announced four funding rounds in five months. In May 2026 it disclosed seed, angel and angel+ rounds totalling nearly RMB 500 million, backed by ZhenFund, HongShan, Meituan Long-Z and Shunwei Capital. ZhenFund had funded the seed round alone in 2025. Shunwei led a Pre-A round in June, followed by a Pre-A+ round in July and a Pre-A++ round in September. Each was described only as several hundred million yuan, and by September cumulative funding had passed RMB 1 billion.