Figure Helix 2.5 Tests Humanoid Autonomy Across 30 Unseen Homes

This Is the Index Test Biped Was Waiting For

When Figure launched Index on August 25, it reported a rapidly growing network of people recording chores and workplace tasks for Helix training. Biped.News treated those upload and contributor numbers as a data-collection story, not proof that the data made the robot better. Human video does not contain the joint torques, gripper forces, or motor commands that a humanoid ultimately needs to act.

The useful question was therefore causal: if Figure holds the downstream robot training and evaluation setup fixed, does pretraining on Index make the resulting robot policy more capable in a new environment? Figure's new Helix 2.5 technical write-up reports such a comparison.

Figure says it trained two policies with identical task-specification data. One started from random weights. The other started from the Index-pretrained Helix 2.5 model. The company says architecture, optimization, hyperparameters, downstream data, and evaluation were held fixed. Under its blind evaluation protocol, the scratch policy completed 9% of zero-shot trials, while the Index-pretrained policy completed 56%.

What "Zero-Shot" Means Here

Zero-shot is easy to overread. Figure is not claiming Helix 2.5 invented a completely new household task after walking through the door. The company first adapted its Index-pretrained foundation model to three defined behaviors using task-specific fine-tuning data collected elsewhere: living-room tidying, towel folding, and bed making.

The zero-shot part applies to the evaluation homes and manipulated objects. Figure says it collected no data in any of the 30 test homes, did not fine-tune or adapt weights there, and did not use evaluation rollouts to choose checkpoints. Each task used one fixed checkpoint across all 30 homes. Evaluation toys, towels, and bedding were held out from task-specification data, with an AI model followed by human review used to check for overlap.

That is a harder setup than a robot repeatedly trained and tested in one apartment. It tests whether the system can cope with changing furniture, bed height, clutter, object appearance, lighting, reach geometry, and walking paths. Figure also made the scoring strict: all toys had to reach the basket, all towels had to be folded and placed correctly, or the entire bed had to meet the predefined completion criteria. Partial progress counted as failure.

Those details make the 56% number more informative than a curated demo reel. They also define its limits. Figure has not published a public benchmark package that an outside lab can replay, and the company remains both the system developer and evaluator.

Helix 02 vs. Helix 2.5: What Actually Changed

The important shift is not a new robot body. It is a different learning recipe. Helix 02 demonstrated long-horizon whole-body control, including logistics work and household manipulation, but those behaviors relied on robot data collected in the environments where the policies operated. Helix 2.5 tries to move more of the learning burden into broad pretraining before the robot reaches the destination.

Figure says Helix 2.5 matched the success rate of a representative Helix 02 policy while using half as much task-specific adaptation data, even though Helix 2.5 was evaluated across 30 unseen homes rather than the environment that supplied its behavior data. That comparison is promising because robot-collected data is expensive. Every hour requires hardware, operators, reset labor, maintenance, and a physical site.

Index changes that economics if ordinary human demonstrations can teach reusable physical representations. Figure's Index network now collects human task video at a scale no robot fleet can match. The new result is evidence that at least some of that diversity can transfer into a humanoid policy instead of remaining a large but indirect video corpus.

The 56% Result Is Encouraging and Still Far From Product Reliability

A 56% strict completion rate can be both a research advance and an unacceptable product experience. That tension is important. Figure's grading is demanding enough that a nearly completed bed or one toy left on the floor counts as failure. Consumers, however, will still notice if a home robot fails close to half of its attempts unless recovery is safe, fast, and mostly invisible.

Figure says Helix 2.5 shows whole-body self-correction, such as backing up to reposition, changing stance, or walking around a bed to repair a fold. Those behaviors matter more than perfection in a short demo because long household tasks create many opportunities for small errors to compound. But Figure has not disclosed the intervention rate, median task time, distribution of failures, hardware fault rate, or how often a person had to reset the environment between attempts.

The company also states that if a human safety intervention was required, the rollout was aborted and counted as a failure. That is the right direction for evaluation, but readers still lack the raw safety-intervention count, near-miss reporting, and standardized home-operation safety evidence needed to compare Helix 2.5 with other home-robot programs.

Figure's Scaling-Law Claim Deserves Attention, Not Automatic Extrapolation

Figure also reports a scaling experiment across four models spanning an eightfold range of Index pretraining data. As data increased, downstream robot-action prediction loss improved smoothly enough that the company says it could forecast the largest run's loss with 0.54% relative error before training that run.

That is technically interesting because robotics has historically lacked the clean scaling curves seen in language models. Physical-world data is expensive, heterogeneous, and tied to embodiment. If Figure can repeatedly predict robot-policy learning from the amount of human-behavior pretraining, it would make data acquisition and compute planning more systematic.

But one internal curve is not a universal law. The forecast concerns model loss, not a guarantee that household task success rises at the same rate. Figure has not published independent replications across different robot bodies, task families, data sources, or organizations. The correct reading is narrower: within Figure's current setup, more Index data produced a smooth enough trend to make a strong held-out prediction.

That still has strategic consequences. Figure says Index is now generating roughly 35 minutes of new human experience every second and that the company has committed $3.5 billion of compute to Helix training. Those are company statements, not proof that scaling alone will produce a reliable household robot, but they show where Figure believes its moat will come from: data diversity plus training infrastructure rather than one hand-designed behavior at a time.

What to Watch Next

The most valuable next step would be external reproducibility. Figure could publish per-task trial counts, confidence intervals, failure categories, intervention statistics, and enough benchmark detail for an outside research group to test the same claims. A larger evaluation across different regions and home types would also show whether the Bay Area sample hides environmental bias.

The second threshold is product reliability. A home robot has to do more than generalize once. It has to work day after day around pets, children, clutter, changing lighting, fragile objects, tight rooms, and people who will not reset a scene for another trial. Success rate, recovery rate, task duration, safety interventions, and remote-assistance burden will matter more than a single headline benchmark.

Finally, watch whether the same pretraining advantage appears in tasks that Figure did not optimize around the Helix 2.5 launch. The strongest evidence for a general foundation model would be a sequence of new behaviors where less robot-specific data is needed each time and the gains survive independent measurement.

Frequently Asked Questions

What is Helix 2.5?

Helix 2.5 is Figure AI's new humanoid neural network, pretrained on the company's Index human-behavior dataset and then adapted to robot tasks. Figure says the model supports whole-body perception, locomotion, manipulation, and recovery on Figure 03.

What does zero-shot mean in Figure's 30-home test?

It means the robot was evaluated in homes and with objects that were not used for environment-specific training or adaptation. The three task types themselves were still specified through fine-tuning data collected elsewhere.

How much did Index pretraining improve success?

Figure reports that an otherwise matched policy trained from scratch completed 9% of strict zero-shot trials, while the Index-pretrained policy completed 56%. The result is company-reported and has not yet been independently replicated.

Does this mean Figure 03 is ready for consumers?

No. The 30-home work is a company-run evaluation, not evidence of a commercial household deployment. Figure has not announced general consumer availability, field-service economics, or production reliability metrics tied to these trials.

The Bottom Line

Helix 2.5 is a meaningful follow-up because it answers the central question left open by Figure's Index launch: whether broad human-video pretraining measurably improves a humanoid's ability to work somewhere new. Figure's controlled result says yes, with strict zero-shot success rising from 9% to 56% across 30 unseen homes while using less task-specific data than an earlier Helix 02 comparison.