Figure AI launches Helix 2.5 for zero-shot home chores

Figure AI has launched its Helix 2.5 model, a major update designed to solve the difficult robotics challenge of generalization in unfamiliar environments. During testing, the model allowed the company's humanoid robots to enter 30 homes they had never previously encountered and immediately perform zero-shot household chores. The robots successfully completed tasks such as tidying living rooms, folding and storing towels, and making beds without prior training on those specific layouts.

This release comes alongside other major developments in the robotics sector. Agility Robotics introduced its Digit 5 humanoid, which safely works alongside humans by slowing down or shutting off its motors when approached. Agility has secured 300 million dollars in multiyear orders, planning a North American launch in early 2027 followed by Europe and the UK later that year. Meanwhile, Tokyo-based MW is building homes with ceiling-mounted robotic arms to handle chores, aiming to sell its first robot-equipped home in 2028. To support mass deployment, UBTECH has built the first plant capable of manufacturing humanoid robots at a 10,000-unit scale in Liuzhou, where robots help assemble other robots and subject each unit to over four hours of testing.

These advancements are driving massive investment. Goldman Sachs now projects that 6.5 million humanoid robots could ship globally by 2035. To support these fleets, Japanese firm GMO has introduced a humanoid ambulance repair van to service broken robots on site. In the startup ecosystem, Mecka AI is reportedly nearing a Sequoia-led funding round at a 500 million dollar valuation, while Travis Kalanick's startup, Atoms, has hired Meta AI veteran Vikas Chandra as its vice president of AI. However, Reuters reports that Boston Dynamics is unlikely to pursue an IPO in 2027 due to unprofitability and a lack of large-scale Atlas deployments.

For AI practitioners, the success of Figure's Helix 2.5 model demonstrates that zero-shot generalization in highly variable, unmapped human environments is becoming commercially viable. Instead of relying on tedious, environment-specific fine-tuning, developers can leverage foundation models that translate directly to real-world deployment. This shift accelerates the timeline for bringing autonomous systems out of laboratories and into dynamic consumer spaces.