Figure AI’s Helix 2.5 Robot Completes Household Chores in 30 Unseen Homes Without Extra Training

A US robotics company says its humanoid robots can walk into a stranger’s home and start doing the washing up — or at least making the bed — without any prior knowledge of the layout, the furniture, or the clutter left on the floor.

Figure AI announced Helix 2.5 on 17 September 2026, posting the news via its @Figure_robot social media account alongside a detailed technical write-up. The company says the neural network completed household chores in 30 rented Bay Area homes it had never seen before, using a fixed software checkpoint with no environment-specific training, fine-tuning, or data collection from those homes.

The robots used in the trial run on Figure’s Figure 03 hardware platform. They were sent into each home and asked to carry out three long-horizon household tasks: making beds, folding towels, and tidying toys and living rooms. None of the 30 homes were part of the original training data.

What the Numbers Actually Show

Across 420 task attempts in those 30 homes, the robots completed 237 successfully. That works out at around 56% — meaning they failed on roughly four in every ten attempts.

Per-task figures reported in specialist robotics coverage suggest bed-making went best, at around 67% success. Towel-folding came in at about 62%. Toy-tidying — the most unstructured of the three — landed at roughly 40%.

That 56% figure looks more striking when set against the alternative. A comparable model trained from scratch on task-specific data, without Figure’s large-scale Index pretraining, managed only about 9% success on the same tests. The Index dataset, built from recordings of human behaviour, appears to be doing a great deal of the heavy lifting.

The @Figure_robot post put it plainly: the robots “arrived with no additional training and started doing useful work.”

How Helix 2.5 Actually Works

Helix is Figure’s family of humanoid control neural networks. The 2.5 version was trained on the Index dataset — a large library of human actions — before being adapted with a much smaller amount of task-specific data. That approach, known as pretraining, is the same broad method used in large language models: train on a huge general dataset first, then fine-tune for specific jobs.

The key claim here is zero-shot generalisation. The same neural network policy, unchanged, was dropped into 30 different homes and expected to perform. No tweaks. No home-by-home adjustments. That’s the part Figure wants people to pay attention to.

Earlier versions of Helix were built around more controlled industrial settings. Helix 2.5 is the first to be tested explicitly in unstructured domestic environments, with all the chaos that implies — different bed sizes, different towel textures, different arrangements of children’s toys across the floor.

What Experts Are Saying

Not everyone is ready to call this a breakthrough. Robotics commentators have pointed out that 56% is a long way from the reliability you’d need before leaving a humanoid unsupervised in your home.

Brett Adcock, Figure AI’s founder and chief executive, has not yet issued a public quote specifically on the 30-home trial results, but the company’s technical post frames the experiment as a proof-of-concept for generalisation rather than a commercial launch.

Independent analysts have been measured. The demonstration shows that zero-shot generalisation across diverse real homes is possible — but they stress that safety, reliability, cost and the sheer complexity of deployment infrastructure remain serious hurdles. A robot that fails four in ten times isn’t ready to look after an elderly person alone, or to be left running in a house with children.

Some critics have also raised concerns about over-reading curated demonstrations. Thirty homes over a short trial period doesn’t capture the full range of edge cases a domestic robot would face over months of daily use.

Price and Availability

Helix 2.5 is a research neural network checkpoint, not a product on sale. The Figure 03 platform is equally unavailable to consumers at present.

Figure AI has reportedly floated a potential consumer target price below the equivalent of around £16,000, based on media reports of the company’s remarks — but that figure is not confirmed by any binding product announcement, and no commercial timeline has been set. Any UK or international pricing remains unverified.

The company has outlined broader plans to scale humanoid systems towards the late 2020s, but specifics are thin.

What This Means for Kent Residents

There are no Figure AI robots operating in Kent, and no trials are planned here. The Helix 2.5 experiment took place entirely in the United States. For residents in the county, the story is one to watch rather than act on — the technology could eventually touch sectors like social care, cleaning and logistics, but the current 56% success rate and the absence of any commercial product mean that day is still some way off. Anyone working in domestic services or care should be aware that this kind of development is moving, but slowly, and UK deployment would need to clear significant regulatory hurdles under Health and Safety Executive rules and the Information Commissioner’s Office data frameworks before it came anywhere near a Kent home.