Latest / Embodied AI 101 / Helix 2.5: The House Changed. The Policy Didn't.
Transcript
- Helix 2.5: The House Changed. The Policy Didn’t. Helix 2.5: Zero-Shot 30-Home Generalization • Figure AI • September 17, 2026. What Figure actually announced. Figure introduced Helix 2.5 with a test of three household behaviors—tidying, folding towels, and making beds—across 30 previously unseen Bay Area homes. The important change is the environment: learned behaviors had to transfer without collecting training data at each destination. Welcome to Embodied AI 101. This episode is about the difference between teaching a robot a chore and teaching it something that survives a change of address. First, a clarification about the accompanying footage. Figure advertised approximately four hours of robot activity across rental homes. That should not be interpreted as a measured, uninterrupted four-hour shift in every home, or as a complete reliability assessment. Video duration and autonomous operating endurance are different measurements. My assessment is that this is meaningful evidence for transfer. Whether that transfer is reliable enough to become a household service is a separate question. Zero-shot has a specific meaning here. “Zero-shot” applies to homes and objects, not tasks. Figure fine-tuned elsewhere, then used one fixed checkpoint per behavior; evaluation results did not determine checkpoint selection. The useful distinction is between deployment generalization and new-skill acquisition. In one case, a previously learned behavior must work somewhere unfamiliar. In the other, the robot must acquire or infer a behavior it was not explicitly trained to perform. A home-generalization result should not automatically receive credit for both. There is important precedent. Physical Intelligence’s April 22, 2025, π0.5 announcement demonstrated mobile manipulators performing chores in previously unseen homes. Its recipe combined robot data from different embodiments and environments with multimodal web data and semantic task annotations. The researchers explicitly framed the advance around transferring existing capabilities to new settings, rather than introducing unprecedented dexterity. That context does not diminish Figure’s result. It changes the question from “Has anyone demonstrated unfamiliar-home operation?” to “What additional embodiment, scope, and evaluation difficulty does this system handle?”. For a professional audience, that is the more productive comparison. “Generalization” needs an accompanying description of what changed, what stayed fixed, and what counted as success. Index is the central technical story. Figure’s proposed source of generalization is Index, its human-behavior data operation. At its August 25 launch, Figure reported more than 16 million uploaded videos and a contributor network spanning over 100 countries. These were collection figures—not a disclosure that every uploaded clip became a training example. The processing pipeline matters more than the headline volume. Figure describes quality filtering, fraud review, deduplication, rebalancing, and annotation. Video embeddings help identify repetition; task quotas and embedding clusters shape the retained distribution; hierarchical captions provide semantic descriptions. My reading is that the potential advantage lies in obtaining useful variation in environments, objects, and execution styles—not simply accumulating more minutes of footage. The Helix 2.5 base was initialized randomly and pretrained on Index, rather than starting from a pretrained vision-language model. That makes the human-to-robot interface especially interesting. What representation carries knowledge from a person manipulating cloth to a humanoid commanding its own joints? Which training targets capture object interaction, and which capture human-specific motion that needs transformation?. Those are questions I would want answered before treating this as a reproducible training recipe. A successful demonstration can establish that a recipe works without yet explaining enough for another laboratory to rebuild it. Why the whole body changes the problem. For architectural context, look back to Helix 02, introduced on January 27, 2026. Figure described a hierarchy in which System 2 handles semantic interpretation and goals, System 1 generates whole-body joint targets at 200 hertz, and System 0 handles balance and contact coordination at one kilohertz. Its low-level controller used retargeted human motion and simulation-based reinforcement learning. These are documented Helix 02 details, not independently verified specifications for every component of Helix 2.5. The engineering motivation is straightforward: locomotion and manipulation constrain each other. Reaching changes balance; moving the base changes reach and visibility; carrying an object constrains how the robot should step. Figure’s earlier architecture explicitly addressed these interactions rather than treating walking and arm motion as unrelated stages. Consider an illustrative bed-making problem. A blanket corner is visible but unreachable from the current stance. The robot needs to reposition without losing its other grasp, avoid trapping the fabric, and obtain a useful view before continuing. I would evaluate that as one coupled behavior—not award separate points for a good grasp and an impressive walk. The operational question is whether the robot can preserve progress while changing how its body interacts with the room. The result that matters—and the unit behind it. Figure reports 9% versus 56% full-task success without and with Index pretraining, holding architecture, task data, downstream training, and evaluation fixed. Safety interventions counted as failures. That controlled comparison is stronger evidence than a montage. It supports pretraining’s contribution within this recipe, not superiority over every alternative dataset or training approach. But success rates need units. Physical Intelligence’s π0.5 evaluation averaged success over individual subtasks, such as moving objects to their correct destinations. That measures task progress differently from requiring an entire chore to finish. Sunday’s July 17, 2026, ACT-2 preview provides another useful comparison. It reports 778 successful autonomous folds and stacks out of 785 garment attempts, or 99.1%, across nine garment types. Its primary success unit is an individual garment, with fold quality evaluated separately. This is not a head-to-head benchmark against Figure’s complete household tasks. Imagine a hypothetical robot that handles almost every object correctly but consistently misses the final item. An object-level metric could look excellent while an all-items-completed metric looks poor. Neither metric is inherently dishonest; they answer different questions. Quality adds another axis. Figure’s rubric separates fold completion from fold neatness. For deployment decisions, I would therefore ask for completion, quality, elapsed time, and assistance burden together. A nearly finished chore may have value—but only if the remaining cleanup does not create more work than the robot removed. A scaling result is not yet a service forecast. Figure’s scaling study used four pretraining data sizes spanning eightfold growth, with model size and downstream training fixed. The forecasted quantity was held-out robot-action prediction loss. That is a useful training-planning signal. The next plot I would want, however, would connect that prediction metric to closed-loop completion, recovery, and safety outcomes. Sunday’s ACT-2 report explicitly examines the relationship between validation loss and real-world success, alongside the effect of data quality. That is a useful methodological comparison: a prediction metric becomes more operationally informative when its relationship to the desired behavior is measured, rather than assumed. There is also a substantial infrastructure commitment behind Figure’s scaling argument. Its September 3 agreement with Nscale described an initial $3.5 billion compute commitment and access to as many as 100,000 GPUs, with initial deployment targeted for the second half of 2027. Those are future capacity plans, not evidence that this experiment already used that infrastructure. My industry interpretation is conditional: if broad pretraining reduces repeated commissioning work at individual customer sites, it could shift engineering effort toward a reusable model. But a large compute commitment is an input to that strategy, not proof of its economics. The commercial test is not simply whether the next model predicts actions better. It is whether each deployed robot requires less human effort to deliver an acceptable result. What would convince an operator?. Robotics builder Jacob Zietek offers a useful perspective in his March 25 essay for a16z. His argument is that deployment requires monitoring, integration, escalation pathways, and customer-focused operations—not just autonomy research. He also argues that useful robotic labor can arrive before perfect autonomy, provided fallback assistance makes operational sense. This is broader practitioner commentary, not an endorsement of Helix 2.5. That avoids two unhelpful extremes: declaring household robotics solved, or declaring every imperfect robot commercially worthless. For the next evaluation, I would ask for four things:. Results organized by home. Show whether failures cluster around particular layouts, furniture, lighting, or object properties. A pooled average can conceal an important deployment boundary. A failure and recovery breakdown. Separate missed grasps, perception errors, timeouts, navigation problems, and safety stops. Measure how often autonomous recovery actually saves the task. An operational ledger. Record setup, charging, resets, supervision, and assistance alongside useful output. Count the human work surrounding the autonomous interval. Externally selected, repeated trials. Test the same frozen system over multiple days, including changed object locations and interruptions. Let an outside evaluator choose at least part of the distribution. These are proposed acceptance tests, not allegations that Figure secretly failed them. I would also resist demanding a single magic reliability threshold. The acceptable boundary depends on the task, the consequences of failure, and the burden of recovery. Sunday’s own proposed evaluation framework usefully separates performance, operating scope, and adaptation cost. Those dimensions should remain visible rather than being compressed into one impressive percentage. The takeaway. Helix 2.5 deserves attention because it makes unfamiliar environments part of the evaluation, rather than merely part of the promotional promise. The next engineering objective should be equally concrete: retain that transfer while reducing unfinished work, improving output quality, and making the cost of intervention explicit. For researchers, the interesting question is how broadly acquired experience becomes transferable control. For operators, it is how much dependable work remains after counting everything needed to support the robot. Generalization expands where a capability can work. Reliability determines whether someone can hand the job over.