Figure's Helix 2.5 tackles chores across 30 unseen homes

Humans can walk into an unfamiliar friend's apartment and immediately make the bed or fold laundry without mapping the room first. Robots, on the other hand, usually freeze if a mattress is two inches higher or a sofa is a different color.

Figure's new neural network model, Helix 2.5, is designed to crack that exact problem. The company sent its humanoid robots into 30 real homes across the San Francisco Bay Area to perform three long-horizon chores completely zero-shot, with no prior mapping, no environment-specific data collection, and no fine-tuning.

Real homes, real clutter, zero partial credit

Figure tested the humanoid across three whole-body tasks: tidying a living room, folding towels, and making a bed. These require a complex mix of whole-body balance, navigating tight spaces, and handling both rigid and deformable items.

The grading was all-or-nothing. To register a success in the living room, the robot had to pick up all 13 to 15 scattered toys and drop them into a basket. For bed making, both pillows and comforter corners had to be aligned and the spread pulled flat.

The robot used whatever couches, beds, and folding surfaces each home already had. None of the specific toys or bedding had appeared in task training data, and a single checkpoint was deployed across all 30 residences.

The robot also showed noticeable whole-body self-correction. When an angle was awkward, it backed up to reset its stance, and it walked around to the far side of a bed to smooth out stubborn folds.

The scaling law of human experience

The magic behind this generalization comes from Index, Figure's global dataset of human behavior. While Figure's earlier Helix 02 started from a pretrained vision-language model, Helix 2.5 was pretrained from scratch directly on Index.

To prove pretraining was driving the leap, Figure ran an ablation test keeping the task data, architecture, and evaluation identical. The policy trained from scratch succeeded in just 9% of trials, while the Index-pretrained model hit a 56% success rate, a more than 6x improvement. Crucially, no single evaluation task accounted for more than 1.9% of the pretraining data.

Helix 2.5 also required half as much task-specific adaptation data as Helix 02, while expanding from single-environment execution to 30 unseen homes.

Figure also measured what it calls the first human-to-humanoid robot transfer scaling law across an 8x data span. Next-action prediction loss dropped predictably enough that Figure forecasted the largest run's test loss to four decimal places before training, with an error margin of just 0.54%. Figure says Index generates roughly 35 minutes of new human experience every second, backed by a $3.5 billion compute commitment.

Why it matters

Deploying humanoids to consumer homes has always hit a wall: nobody wants a robot that requires engineers to map and fine-tune their living room for days. Helix 2.5 shows that zero-shot whole-body generalization is actually feasible when grounded in broad physical experience.

That said, you should not expect to buy an autonomous home butler anytime soon. While a 56% success rate on strict, multi-step domestic chores is a scientific triumph, a 44% failure rate in a real household is still a mess waiting to happen.

In our view, Figure's approach is brilliant because it finally imports the true LLM playbook into physical robotics: broad unsupervised pretraining followed by light downstream behavioral fine-tuning. By proving that human-to-robot transfer scales predictably with compute and data, Figure turns robot capability into an engineering problem rather than a collection of fragile edge-case fixes.

The downside is the extreme capital barrier. With a $3.5 billion compute pipeline and massive proprietary data collection, this roadmap threatens to turn embodied AI into a playground exclusive to deep-pocketed tech giants.