Figure's $3.5B Data Bet Pays Off as Helix 2.5 Zero-Shot Generalization Hits 56%
Figure has introduced Helix 2.5, a humanoid control model that completed three whole-body household tasks in 30 Bay Area homes it had never entered, without collecting data in those properties and without fine-tuning on their layouts, furniture, or objects. In blind evaluations, Figure puts Helix 2.5 zero-shot generalization at 56% on full-task trials, against 9% for a model trained from scratch on the same tasks. The release, announced September 17, is a software update that runs on Figure's existing Figure 03 platform rather than a new hardware generation.
The three tasks are long-horizon and whole-body: tidying a living room, folding towels, and making a bed. Figure rented the homes and reset the rooms between trials, then sent in a robot with no prior exposure to those spaces or to the objects it would handle. No single evaluation task accounted for more than 1.90% of the model's total pretraining data.
Helix 2.5 Zero-Shot Generalization: Why the 9% to 56% Gap Is the Story
The test setup is harder to game than a lab demonstration. Renting 30 separate homes and resetting each room between trials costs more than building one staged set, and it removes the chance that the model is recognizing a fixed environment it has already seen. Every room the robot entered was a room it had never occupied.
Figure then ran two models through the same blind test. One learned the three tasks from scratch. The other learned them on top of Index, Figure's dataset of recorded human behavior, after pretraining from random initialization. The 47-point spread between the runs isolates a single variable: what the model saw before it touched the task.
Hardware, sensors, and actuator limits were identical in both conditions, which makes the jump hard to attribute to mechanical improvement. Figure also reports that the Index-pretrained model needed 50% less task-specific data than the prior Helix 02 version while expanding operational scope roughly 30x.
The 1.90% ceiling carries weight for anyone assessing the claim. Each of the three evaluation tasks is a rounding error inside the pretraining corpus, so the performance came from the breadth of Index rather than from targeted exposure to the tasks themselves.
The three behaviors were chosen to stress different parts of the policy. Folding towels is deformable-object manipulation, where the target state shifts with every fold and there is no fixed pose to snap to. Making a bed demands whole-body coordination across a large area, with the robot moving its base while keeping contact with fabric. Tidying a living room combines object recognition, grasp selection, and placement decisions that depend on the room itself.
The 50% reduction in task-specific data matters for a separate reason. Task-specific data is the expensive kind: it needs a robot, an operator, and a physical setup for every new skill. Halving that requirement lowers the marginal cost of teaching Helix 2.5 a fourth or fifth household behavior, which is the step that turns a three-task demonstration into a product catalog.
Index: Roughly 2,100 Hours of Human Experience per Hour
Index now generates about 35 minutes of new human experience data every second, which works out to roughly 2,100 hours of captured behavior every hour, or about 50,000 hours a day. Figure describes a human-to-humanoid transfer scaling law, in which doubling pretraining data produces predictable gains in robot action prediction. The company has committed $3.5 billion in compute to training Helix.
Human data also sidesteps a hardware bottleneck. Collecting robot demonstrations requires a working robot for every hour captured, which caps throughput at fleet size. Index decouples capture from the humanoid, so the corpus grows independently of how many Figure 03 units exist.
| Metric | Figure Helix 2.5 result |
|---|---|
| Zero-shot full-task success, blind test | 56%, against 9% trained from scratch |
| Homes evaluated with no local data collected | 30, San Francisco Bay Area |
| Task-specific data versus Helix 02 | 50% less |
| Operational scope versus Helix 02 | About 30x |
| Largest single evaluation task share of pretraining data | Under 1.90% |
| Index data generation rate | About 35 minutes per second |
| Compute committed to training Helix | $3.5 billion |
A predictable scaling relationship changes planning. If returns on data are measurable instead of erratic, data collection becomes a capital line item with a forecast attached, in the way language model labs now treat token supply. Figure's position is that the binding constraint sits on the data side, and the $3.5 billion is the compute budget that consumes it.
Compute demand then tracks the corpus rather than staying flat, which is one reason the commitment is counted in billions of dollars instead of hardware units. A larger Index requires proportionally more training capacity to convert into better policies, so the two budgets rise together.
The open question is embodiment mismatch. Index records human bodies, and the policy drives a machine with different joints, reach, and grip. Figure's scaling law is a claim that the gap narrows predictably as the corpus grows; the 56% result is evidence for that claim rather than proof that the gap closes entirely.
What 56% Still Leaves Unfinished
Read the Helix 2.5 zero-shot generalization figure the other way and roughly two in five long-horizon attempts fail in a home the robot has never seen. For a research result that is a wide step forward. For a commercial service it defines how much human supervision the product requires, since a robot that fails 44% of full-task trials cannot be left alone in a paying customer's house.
The economics are unforgiving at current rates. Commercial humanoid cleaning in San Francisco has carried rates near $30 an hour. A 44% failure rate on multi-step chores converts a large share of billable time into rework unless the operator narrows the task set to the behaviors with the highest success probability.
Task scoping is the practical lever. Tidying a living room, folding towels, and making a bed each tolerate error differently, and an offering built around the single highest-yield behavior can lift effective reliability above the headline number without any change to the model.
That scoping decision also points to the nearer buyer. The 30-home evaluation covers a fixed set of three tasks rather than open-ended housework, which suits a commercial service with a defined checklist before it suits a consumer product expected to handle any room.
Where Figure Sits Against the Field
Foundation models for robot control are a crowded category. Nvidia's GR00T N-series, Google's Gemini Robotics models, and the π family of policies all pair a vision-language module that reads the scene with a separate module that generates motor commands. Helix uses a related split, combining an 80-million-parameter control transformer with a slower reasoning component.
The Figure 03 platform provides the physical layer: whole-body locomanipulation, active perception, bimanual coordination, and autonomous error recovery. Helix 2.5 is the control layer on top of it.
What separates Figure's claim is where the differentiation sits. Accelerators can be bought by any well-funded competitor, and several have bought them. A corpus of human behavior arriving at 35 minutes per second cannot be purchased on the open market. That asymmetry, more than any single benchmark score, is the case for the $3.5 billion.
Figure CEO Brett Adcock has described zero-shot generalization as the industry's most sought-after capability, the point at which a robot can enter an unfamiliar building and begin useful work without a setup phase.
Why this matters
Helix 2.5 shifts the humanoid competition from who builds the best body to who controls the best behavior data, and Figure has priced that shift at $3.5 billion in compute. For operators, the 56% figure sets the supervision cost of any near-term deployment; for investors, it makes data pipelines rather than actuator supply chains the asset worth tracking. Index keeps adding about 35 minutes of human experience every second, so the next point on the transfer scaling curve is already being recorded.
Sources
📰 Hacker News AI Digest 2026-09-18 · Issue #286 · ghub1821239/agents-radar
Artificial Intelligence · fmhy/FMHY Wiki - AI Chatbots
Related Articles
- The Phase-Aware Inflection: Predictive Tech Moves Past Static Averages
- Datacurve DeepSWE Benchmark Identifies Major Errors in AI Coding Tests
- OpenEvidence Rejects $20B, Betting on Specialized AI
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.