Figure AI has demonstrated a significant breakthrough in robot generalization by deploying its humanoid robots into 30 unfamiliar homes where they successfully completed household chores without any prior training in those specific environments. The company's new Helix 2.5 AI control model achieved a 56% success rate across three tasks, compared to just 9% for a comparable model trained from scratch, marking what CEO Brett Adcock described as the most important project Figure has undertaken.

What Makes This Different From Previous Robot Demonstrations?

For years, robotics companies have released polished videos of humanoids folding shirts or sorting items in pristine laboratory settings. What rarely gets mentioned is the engineering effort behind the scenes: engineers spend weeks mapping rooms down to the centimeter, locking in fixed lighting, and fine-tuning models to specific countertops. The moment you change the layout or swap the towels, the routine falls apart.

Figure's approach bypasses this site-specific training entirely. The company rented 30 residential properties across the San Francisco Bay Area and tested its humanoid robots on three whole-body household tasks without collecting any training data inside those homes. The objects used in testing, including toys, towels, and bedding, never appeared in the task-specific training data.

The three tasks tested were straightforward but demanding: making beds by positioning pillows and smoothing comforters, folding towels and placing them in baskets, and tidying living rooms by picking up 13 to 15 scattered toys. Success required complete task execution with zero partial credit.

How Did Figure Achieve This Generalization Breakthrough?

The key to Helix 2.5's performance lies in a dataset called Index, which Figure publicly launched on August 25, 2026. Index is built from large-scale recordings of human behavior and serves as a pretraining foundation for the robot's neural network. Rather than training the model directly on robot demonstrations in specific homes, Figure first trained it on broad human activity data, giving the system an intuitive understanding of how people interact with household objects and spaces.

The performance difference was stark. When Figure tested a control policy trained from scratch without Index pretraining, it achieved a zero-shot success rate of just 9% across the homes. When running the Index-pretrained Helix 2.5 model, that success rate climbed to 56%.

Figure also demonstrated what it calls a human-to-robot transfer scaling law. By training four models across an eightfold expansion in pretraining data while keeping downstream tuning fixed, action prediction error dropped predictably with every doubling of data. The company claims the trend was consistent enough to forecast the validation loss of its largest model run to four decimal places prior to training.

Breaking Down the Task-Specific Results

Performance varied across the three household tasks tested. Bed making proved the easiest for the robot, succeeding in 94 of 140 attempts for a 67% success rate. Towel folding came in second at 62%, with 87 successful completions out of 140 trials. Tidying toys was the most challenging task, landing at 40% with 56 successful runs out of 140 attempts.

The hardware running all of this is Figure's 03 platform, the company's current generation of humanoid robot. Helix 2.5 is the software layer handling perception, decision-making, and physical control simultaneously.

What Are the Real-World Limitations?

While a 56% success rate represents a genuine achievement in zero-shot generalization, it also reveals the gap between controlled testing and everyday household reality. The robot fails more than four times out of ten, which would be unacceptable for a household appliance you'd actually use on a regular basis.

The evaluation homes themselves offer important context. Figure conducted tests in holiday rentals and Airbnbs rather than occupied family homes. These spaces are clean, well-lit, and vacant, with company engineers hovering within arm's reach of the robot as it navigates around furniture. An active Australian family household, by contrast, is far more chaotic: dogs running through kitchens, children leaving school bags across doorways, and family members walking past.

Navigating static furniture in an empty rental without prior mapping is a genuine achievement in autonomous spatial reasoning. However, safe, dynamic coexistence with people remains an unresolved challenge.

How Does Figure's Approach Compare to Competitors?

While Figure focuses on human video pretraining for humanoid platforms, other companies like Physical Intelligence are tackling generalization from complementary angles. Physical Intelligence's pi 0.5 model uses heterogeneous co-training, consuming a mix of multimodal web data like image captioning and visual question answering, combined with action data from various robot platforms.

Physical Intelligence evaluated pi 0.5 across unseen homes on domestic tasks including clearing dirty dishes into sinks, loading clothes into hampers, and wiping bench surfaces with sponges. Their ablation testing revealed that web-scale multimodal data was the single biggest contributor to helping the robot recognize unfamiliar household items.

Steps to Understanding Figure's Technical Roadmap

• Foundation Model Pretraining: Figure trained Helix 2.5 on Index, a large-scale dataset of human behavior recordings, before any task-specific training occurred in target environments.
• Zero-Shot Evaluation: The robot received no additional training specific to the 30 test homes or the objects inside them, testing true generalization capability.
• Scaling Law Validation: Figure demonstrated that robot action prediction improves predictably with each doubling of pretraining data, suggesting that more data and compute will translate directly into better performance.
• Compute Commitment: The company has committed $3.5 billion in computing resources to training Helix, with Index now generating about 35 minutes of human-experience data every second.

Helix 2.5 also represents an efficiency gain over its predecessor. The model cut the amount of task-specific data needed to specify a behavior in half while expanding its scope from a single environment to 30 unseen homes. Figure describes that as making behavior specification 2x cheaper while increasing deployment scope 30x.

"The holy grail for robotics is being able to generalize: doing work in unseen places. We rented 30 homes in the Bay Area and are doing tasks without any new training. Helix 2.5 was built to answer a harder question: can a humanoid enter a home it has never seen and immediately get to work, with its whole body, on its own?" said Brett Adcock, founder and CEO of Figure.

Brett Adcock, Founder and CEO at Figure AI

What Does This Mean for the Timeline to Consumer Robots?

These announcements confirm that robotic foundation models are progressing rapidly, but the timeline remains measured. We are likely still years away from walking into a retail store to purchase a domestic humanoid for around $15,000 to handle Saturday morning cleaning.

The value of these technical updates lies in validating the software roadmap. For years, the industry debated whether physical manipulation could benefit from the foundation model scaling laws that drove large language models. The data from Figure and Physical Intelligence proves that physical intelligence scales with data.

Training foundation models on broad human activity and multimodal data gives machines a basic intuitive understanding of the physical world before entering a room. The era of hand-coding every single room is finally coming to an end, though the journey toward reliable, everyday household robots remains in its early stages.