Figure’s Helix 2.5 Shows Zero‑Shot Success in 30 Unseen Homes

On September 17, 2026, Figure announced the launch of Helix 2.5, a brand‑new control architecture designed for humanoid service robots. The company’s official press release emphasized a field experiment where a single instance of the model was taken to thirty distinct residences that had never appeared in any of the training datasets.

During the experiment the robot was asked to carry out three routine household chores: it had to tidy up a living‑room area, fold a stack of towels, and make a bed, all using the furniture, fixtures and objects that were already present in each home. Importantly, the team performed no prior mapping of the spaces nor any cataloguing of the items before the robot entered the houses.

Strict Separation of Training and Evaluation Data

Figure underscored that none of the data collected from the evaluation homes was ever fed back into the training pipeline. Independent human auditors, working alongside the company’s engineers, confirmed that none of the objects, room layouts or spatial relationships observed during the test were included in the specification data that had been used to train Helix 2.5.

The experiment relied on a single control checkpoint of the model that was applied uniformly across all thirty environments. The neural‑network weights were left untouched—no fine‑tuning took place—and no selection mechanisms based on early performance were employed.

Performance Compared With a Target‑Specific Policy

Figure reported that Helix 2.5 matched the success rate of Helix 02, an earlier version that had been trained directly inside the target environment. The crucial distinction, according to the company, was that Helix 2.5 achieved this parity while consuming roughly half the amount of adaptation data required by Helix 02.

The system also demonstrated on‑the‑fly physical corrections. For example, when a planned motion would have struck a bed frame, the robot automatically backed away, adjusted its posture, or navigated around the obstacle before continuing with the folding sequence.

Independent Assessment by ABC Tecnología

ABC Tecnología published an independent assessment that recorded a raw success rate of 56 % across the thirty homes. The outlet also quoted local robotics experts who warned that a failure rate approaching one in two would be far too high for a consumer‑grade household assistant.

Figure replied that the purpose of the test was to evaluate generalisation rather than to demonstrate full domestic autonomy. Only the three pre‑specified tasks were measured, and the criteria for success had been defined in advance of the experiment.

• 30 previously unseen homes
• 3 household chores (tidying, towel folding, bed making)
• Single model checkpoint, no weight adaptation
• 56 % overall success rate reported by ABC
• Half the adaptation data compared with Helix 02

The company also disclosed the scale of its computational investment: the Index platform that powers Helix generated roughly 35 minutes of new human‑level experience per second, and a total of $3.5 billion in compute resources were allocated to train the model.

While the results clearly demonstrate measurable zero‑shot generalisation, they do not imply that a robot can now operate fully autonomously in any domestic setting. The test was deliberately limited to a narrow set of tasks and predefined success metrics.

For English‑speaking organisations considering the deployment of service robots, the findings suggest two practical takeaways. First, a single, well‑trained model can cope with a variety of previously unseen environments without the need for costly on‑site retraining, potentially cutting rollout time and expense. Second, the current success rate indicates that human supervision or fallback mechanisms remain essential to ensure reliable operation in real homes.

In summary, Figure’s Helix 2.5 showcases a promising step toward zero‑shot domestic robotics, yet the modest success ratio and the constrained experimental scope remind developers that widespread, unsupervised household use remains a future challenge.