- Former Alibaba executive raises mega-funding round to develop world models for robots
- Hangzhou startup bets on first-person ‘Ego’ data to help robots learn from real-world experience
Hangzhou-based embodied AI company Infiforce (原力无限) has raised nearly 1 billion yuan ($148 million) in its Series A and A+ rounds, as competition in robotics shifts from hardware toward machine intelligence.
The company announced the funding on August 14. Dunhong Asset Management and leading state-backed investors led the rounds, with Zhejiang University Science and Technology Innovation Group, Yandu State-Owned Capital and Lishui state-owned investors participating. Existing investor CCV also increased its stake.
Founded in 2023, Infiforce is led by founder and CEO Bai Huiyuan (白惠源), a former Alibaba vice president who spent more than 15 years at the company, overseeing businesses including Alibaba China, Tmall, DingTalk, Alibaba Cloud and Alibaba Health.
The company plans to use the new funding to develop its AtomBrain causal world model, upgrade its AI infrastructure platform DataGrid and scale delivery of robots in multiple form factors.
Betting on first-person data
Infiforce’s key technical bet is “Ego-native” data: collecting real-world human interaction experiences from a first-person perspective rather than relying mainly on third-person demonstrations.
The company says its AtomVLA model achieved a 97% success rate on the LIBERO benchmark, while its HiMem-WAM world model reached 97.7%, outperforming π0.5 by about 22.5%.
Infiforce says it will soon release what it describes as the world’s first embodied model built around high-quality Ego data.
The company is already testing its technology in more than 100 real-world scenarios across over 30 Chinese cities, spanning commercial services, logistics and manufacturing.
The bigger bet is that robotics will increasingly be won by learning rather than hardware alone. By combining first-person data, world models and physical robots, Infiforce is seeking to build a closed loop in which robots continuously learn from the physical world.
If that approach works, the competition in embodied AI could shift from who can make robots move to who can make them learn faster.
