Generalization

A policy that cleans the kitchen it was trained in has demonstrated control. A policy that cleans a kitchen it has never seen has demonstrated something much rarer, and the distance between the two is the subject of this module. The training distribution of robot data is narrow: demonstrations come from a small number of labs, with specific robots, cameras, and lighting. Real homes bring different geometries, object instances, backgrounds, and clutter, and physical interaction punishes small perceptual errors in a way text generation never does: a one-centimeter grasp error drops the object. Closing that distribution shiftdistribution shiftThe mismatch between the data distribution a model is trained on and the distribution it must act on at deployment. Robot demonstrations come from a small number of labs with specific robots, cameras, and lighting, while deployment spans unseen homes, objects, and clutter, so a policy can score well on held-out data from its own training distribution and still fail in a new kitchen. Open-world evaluations measure performance under exactly this gap: pi0.5 was evaluated in three real homes absent from its training set, and Lin and colleagues measured how generalization to new environments and objects scales with the diversity of the training data rather than its raw volume., without collecting new data at every site, is the generalization problem. It is the gap between a demo and a product, and it is where the strongest systems of 2026 are now being measured.

The strongest quantitative argument that the gap closes with data is a scaling law, and it is the thing to look at first. Drag the extrapolation horizon slider past the measured range and the two series continue as dashed lines inside a band that brackets the law holding against a plateau. Where you land depends entirely on which of those two you believe, which is the argument the rest of this module sets up.

What pi0.5 demonstrated

The π0.5 result of April 2025 is the reference point for open-world manipulation: a single end-to-end vision-language-action modelvision-language-action modelA robot policy built by adapting a vision-language model, pretrained on web-scale image and text data, to output robot actions. RT-2 established the approach by representing actions as text tokens and co-fine-tuning on web and robot data together, showing that semantic knowledge from web pretraining transfers into control. The same recipe underlies OpenVLA and the π series of generalist policies. cleaning kitchens and bedrooms in homes that were entirely absent from its training data Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025. The evaluation behind that claim is worth reading precisely, because its shape defines what the result does and does not say. The final evaluation ran in three real homes, three kitchens and three bedrooms, with ten trials per task on chores like clearing dishes into a sink and packing items into a drawer, each episode spanning minutes Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025.

The training recipe explains where the generality comes from. Only about 400 hours of the mixture is mobile-manipulator data collected in real homes; 97.6% of the training examples in the first training phase come from somewhere else, meaning other robots, laboratory task data, web vision-language data, and semantic subtask prediction Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025. The ablations show this transfer is essential rather than incremental: strip the non-action data and the open-world result disappears. The most informative control in the paper is a model trained directly on data from the test homes. It scores about the same as the 104-location model that never saw them, the strongest published evidence that broad co-training closes the generalization gap on this task set Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025.

One distinction does a lot of work here. Web-scale vision-language pretraining buys semantic generalization: the model knows what "the spatula" is even in a new kitchen. It does not by itself buy motor generalization: how to grasp that spatula, with this gripper, in this geometry. The first comes free with the backbone; the second is what the robot data is for, and it is the harder half of the problem.

What pi0.7 adds

π0.7 (April 2026) pushes on two axes beyond π0.5, and the architecture behind it is covered in The Pi Line; what matters here is the shape of the claims.

The first is compositional generalization: recombining trained skills into tasks that were never trained. The headline example is loading a sweet potato into an air fryer, an appliance task with no robot demonstrations in the training set. Physical Intelligence's own account of the zero-shot attempt is the honest one: the model "makes a reasonable attempt, performing part of the task after a few false starts, but not finishing it fully" Ai 2026↓π0.7: a Steerable Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. With step-by-step language coaching the robot completes the task, and the coaching transcripts then fine-tune a high-level policy that runs the task autonomously, with no additional teleoperationteleoperationDirect remote control of a robot by a human operator, used in robot learning mostly to record expert demonstrations. The ALOHA rig made the setup cheap and precise: the operator back-drives a leader arm that is a joint-for-joint twin of the follower the policy will run on, so the recorded motion is already in the robot's own action space. Mobile ALOHA extended the same leader-follower scheme to a mobile base for whole-body tasks. data collected Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Short-horizon compositional tasks do work outright: pressing a french-press plunger, scooping rice into a cooker.

The second is cross-embodimentcross-embodimentLearning from, or transferring to, robots with different morphologies: different joint counts, action dimensions, and sensor suites. The Open X-Embodiment project pooled data from 22 robot embodiments into a single training mixture and showed that the resulting RT-X policies carry skills across platforms. The gain is not free, because each body defines its actions differently, and the field answers that mismatch with several incompatible conventions. transfer. π0.7 folds laundry on a bimanual UR5e, a robot that was never trained on any laundry-folding task, and reaches 85.6% task progress and 80% success. The comparison baseline is a panel of expert teleoperators attempting the same task on the UR5e for the first time, operators ranked in the top 2% by experience with a mean of 375 hours of it; they managed 90.9% progress and 80.6% success Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The transfer is not mimicry. On the bag-packing task, teleoperators on the source robot hold the bag open with one arm while inserting with the other; π0.7 on the taller UR5e instead discovers a single-arm pick-and-place suited to its reach Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026.

Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026 DeepMind reports a complementary result from the same announcement: its Gemini Robotics On-Device 2 model adapts to new bi-arm robot embodiments with just a few hours of adaptation time, typically with less than 200 examples, a vendor-reported figure.

What these results do not demonstrate

The limits are as much a part of the papers as the results, and they deserve the same plain statement.

First, the air fryer attempt does not finish the task without coaching. Compositional generalization in π0.7 is emergent but not reliable: the authors themselves describe only "early signs of compositional task generalization," and the completed-task videos are coached or post-trained on coaching transcripts Ai 2026↓π0.7: a Steerable Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. A capability that needs a human walking the robot through the task is real, and it is not autonomy.

Second, nothing here demonstrates high reliability across the full distribution of homes. π0.5's headline evaluation is three homes and ten trials per task; that is evidence of a genuine capability, not of a reliable one, and the difference between the two is the subject of The Reliability Gap and the measurement problems in The Evaluation Crisis. No published result shows a generalist policy sustaining better than 95% successsuccess rateThe standard headline metric of robot learning evaluation: the fraction of attempted episodes in which the policy completes the task. Trial count, time limit and scene distribution belong with the reported rate. Toyota Research Institute's Large Behavior Model study analyzed 1,800 real-world rollouts and warned of a significant risk of statistical noise from insufficient power, not a measured prevalence of incorrect papers. An inconclusive comparison does not establish policy equivalence. across a broad task distribution in unseen environments.

Third, semantic generalization is not motor generalization, and the evaluations lean semantic. Following an instruction about an unseen object in a seen category is a different measurement from manipulating an object whose physics are new, which is where the tactile and contact problems covered in Dexterity take over.

Fourth, transfer has a shape. It works best when source and target share task structure, grasp, move, place, and the π0.7 paper's own ablations show performance degrading as the embodiment gap widens before the latest model recovers it Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Transfer across fundamentally different morphologies, a parallel gripper to a dexterous hand, remains limited.

The scaling evidence, and what it measures

The strongest argument that the open-world gap will close with data is a scaling law. NVIDIA's EgoScale trained a vision-language-action model on 20,854 hours of action-labeled egocentric human video, more than twenty times prior efforts, and found that held-out human-video validation loss follows a clean log-linear law in data scale, $L = 0.024 - 0.003 \ln D$ with $D$ in thousands of hours, fit over five scales from 1k to 20k hours at $R^2 = 0.9983$ Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. The law is not only an offline metric: downstream robot performance after post-training rises monotonically with pretraining scale, from 0.30 average task completion at 1k hours to 0.71 at 20k, and the full recipe improves average success rate by 54% over a no-pretraining baseline on a 22-DoF dexterous hand Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. The same corpus is already load-bearing elsewhere: 20K hours of EgoScale video enter GR00T N1.7 pretraining directly through the shared relative end-effector space described in Cross-Embodiment Transfer NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026.

The chart at the top of this module plots that law over its measured range and lets you do the thing the paper declines to do: extrapolate. Everything right of 20k hours is inference, which is why the dashed band is drawn as wide as the disagreement it represents.

Three properties of the chart deserve attention. The measured range ends at 20k hours; everything right of the boundary is inference, and the paper's authors say so explicitly. The band is a pair of scenarios, not a confidence interval: nobody has measured the variance of a 100k-hour run. And the completion fit crosses 100% near 250k hours, which is impossible, so the curve must bend somewhere; the only question the chart cannot answer is where. Above all, the law is fit to validation loss, a proxy that correlates with real-robot performance but is not a success ratesuccess rateThe standard headline metric of robot learning evaluation: the fraction of attempted episodes in which the policy completes the task. Trial count, time limit and scene distribution belong with the reported rate. Toyota Research Institute's Large Behavior Model study analyzed 1,800 real-world rollouts and warned of a significant risk of statistical noise from insufficient power, not a measured prevalence of incorrect papers. An inconclusive comparison does not establish policy equivalence. measured across unseen environments. No one has published a scaling law for that metric.

Whether the proxy law is enough is a live, named disagreement. The scaling camp points to exactly this chart: EgoScale's law with no saturation in the explored regime, π0.7's compositional generalization from a deliberately diverse mixture, and investor analysis from Bessemer arguing that scaling laws are beginning to show up in robotics data and that the ChatGPT moment for robotics is "not years away", though in their words "it's not here yet" Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026 Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. The structure camp answers that robotics has no internet-scale physical data source and never will: Ken Goldberg frames the shortfall as a 100,000-year data gap and argues good old-fashioned engineering, not pure scaling, closes it Goldberg 2025↓Good old-fashioned engineering can close the 100,000-year "data gap" in roboticsKen Goldberg, Science Robotics, 2025. Rodney Brooks argues the missing ingredient is tactile sensing and engineered priors, not more video Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025, and Karcini and co-authors argue policy scaling alone misses the supervision layer entirely: converting unstructured physical experience into grounded robot supervision Karcini 2026↓Robots Need More than VLA and World ModelsElis Karcini, Faisal Mehrban, Quang Nguyen et al., 2026. The observation that would settle it is precise: a 10x or 100x increase in data, with no architectural change, producing proportional gains in real-world success rate across unseen environments, or failing to.

What solved would look like

Generalization is solved when a single policy achieves better than 90% success across many unseen homes, dozens rather than three, with no per-site data collection or fine-tuning, over a broad distribution of tasks, objects, and configurations, reported in a published and reproducible evaluation. No current system has been evaluated this way. The strongest open-world result on record covers three homes with ten trials per task, and the strongest scaling evidence measures a proxy loss, not real-world success.

The chart above makes the distance concrete. The solved bar sits at 0.90 on the completion axis; the best measured point is 0.71, on a friendlier metric, after post-training, on tasks the evaluators chose. The optimistic extrapolation crosses the bar only near 111k hours of human video, five times past the end of the measured range, deep inside the region where the honest answer is that we do not know. That is the open-world gap in one picture: the field can now measure the slope of progress, and still cannot tell you how far the slope carries.