GEN-0 vs HumanScale: Robot Data vs Human Data at Scale

GEN-0 is introduced in a Generalist AI blog post as a new class of embodied foundation models built for multimodal training directly on raw physical interaction. The post lists several headline claims: a phase transition at 7B parameters, where smaller models stop absorbing new data and larger ones keep improving; scaling laws that link more pretraining data and compute to better downstream post-training performance; an approach called Harmonic Reasoning that interleaves asynchronous streams of sensing and acting tokens; and tests on 6DoF, 7DoF and 16+DoF semi-humanoid robots.

It is a company announcement rather than a peer-reviewed paper. Its figures come from the post’s own charts and text. It does not state whether the models or the dataset are available outside the company, so this page treats availability as not stated.

HumanScale is an arXiv paper, “HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining” (arXiv 2606.20521). It runs a controlled comparison: the same autoregressive world-action model is pretrained on either egocentric human video or teleoperated real-robot data at matched hours, then post-trained on the same set of real-robot tasks, a 15-task, 1,500-trajectory set drawn from AgiBot World, and evaluated on Seen and Unseen splits.

The abstract reports that, with the same amount of pretraining data, models pretrained on egocentric data reach a 24% lower validation loss on real-robot action prediction and 52.5% and 90% higher success rates on in-distribution and out-of-distribution real-robot tasks. The authors describe their results as encouraging but still preliminary.

GEN-0’s post describes an in-house robotics dataset of more than 270,000 hours of real-world manipulation data and refers to its “robot data operations” and to “1,000s of data collection devices and robots”. It does not say how much of the 270,000 hours is robot teleoperation, how much comes from other devices, or whether any of it is human video.

HumanScale is explicit. Its two pretraining arms are egocentric human video, a 5,000-hour subset of the egocentric portion of HumanNet selected for diversity in objects, skills and environments, and a 5,000-hour multi-embodiment real-robot corpus. The paper frames the choice as which source supplies coverage at scale for pretraining, with post-training handling embodiment alignment, rather than as one modality being intrinsically better.

So the robot-versus-human framing fits HumanScale directly and GEN-0 only loosely. The two are not arms of one experiment: neither source evaluates the other.

GEN-0’s post describes the infrastructure around collection more than the collection itself: custom hardware, dataloaders and network infrastructure, including new dedicated internet lines, to carry uplink bandwidth from data collection sites worldwide. It also notes that its pretraining and post-training datasets were collected by different people in different environments, to keep evaluations valid. Sensor suites, camera placement and label methods are not stated.

HumanScale’s human video is curated rather than captured for the study. The paper describes egocentric footage as typically captured passively by an off-the-shelf camera or headset, or harvested from the web, and notes that the remaining difficulty is curation: filtering, deduplication, privacy review and pseudo-action labeling through hand-pose retargeting. Its Table 1 lists HumanNet as web curation. The robot arm of the study is aggregated from existing real-robot datasets.

On scale, GEN-0 reports over 270,000 hours and more than 10,000 new hours per week, and a figure in the post states it is trained on orders of magnitude more real-world manipulation data than some of the largest robotics datasets as of November 2025. HumanScale trains on 5,000 hours per arm, limited by available real-robot data.

On scaling behaviour, GEN-0 reports a power-law relationship between pretraining dataset size, counted in action trajectories, and downstream validation error after post-training, measured across 16 task sets. It gives the form L(D) = (D_c / D)^α and says the fit can estimate how much pretraining data a target error needs. It also reports from large-scale ablations that data quality and diversity matter more than sheer volume.

HumanScale reports that as egocentric pretraining grows from 100 to 5,000 hours, best post-training action loss falls monotonically, from 0.0234 to 0.0204 on unseen tasks, with log-linear fits of R² 0.86 (seen) and 0.94 (unseen). On unseen tasks, real-robot pretraining stays near 0.025 across scales, reaching 0.0254 at 5,000 hours. The axes, models and metrics differ from GEN-0’s, so read the two curves separately.

GEN-0 reports that scaling trends carry into real-robot performance. In blind A/B evaluations, more pretraining data raised downstream task success for models post-trained on only 5.6 hours (about 1%) of task-specific data, with up to 99% peak success in some cases when large-scale pretraining was paired with all 550+ hours of task data. The post states pretraining and post-training data did not overlap. It says the architecture works across robots and was tested on 6DoF, 7DoF and 16+DoF semi-humanoid robots.

HumanScale evaluates on a single real-robot platform, an AgiBot bimanual robot, across three tasks: placing a cup, sorting fruits and vegetables, and stamping. The egocentric-pretrained model reaches 92.5% in-distribution and 90.0% with unseen objects, against 40.0% and 0.0% for a Wan2.2 baseline with no embodied pretraining. For embodiment, the paper’s position is that egocentric video’s main weakness, the embodiment gap, is what post-training with a smaller amount of kinematically aligned robot data corrects.

For GEN-0, the post leaves open how the 270,000 hours were captured and labeled, how they divide by source, what the model and dataset availability is, and what its data-mixture experiments, shown only as anonymized partner and class results, imply for any single capture method.

For HumanScale, the paper itself lists limits. Training is capped at 5,000 hours because real-robot data is limited. Evaluation is mainly on one world-action model built on Wan 2.2, and a parallel study with vision-language-action models is left to future work. Real-robot rollouts use one platform and three tasks, and their baseline has no embodied pretraining. The actions on human video are pseudo-labels from retargeted hand pose, and matching by hours does not match trajectory counts. The paper does not state commercial licence terms for the data in the text read for this page.

This page is an independent reading of two public documents, and Firsthand is a separate data provider with no role in either result. Custom egocentric collection does not replicate GEN-0’s dataset or HumanScale’s experiment, and it does not provide robot teleoperation data. What it offers is a way to fill the coverage a buyer’s own pretraining or mid-training mix lacks, with human first-person demonstrations captured to the buyer’s own spec.

A collection starts with a written skill spec naming the scenes, tasks, objects and people the corpus is missing, usually proved by a Spec Pilot, a small first batch, before volume scales. Capture uses a head camera and two wrist cameras hardware-triggered on one PTP clock, and batches ship with 21-joint 3D hand pose with per-joint visibility flags and action segments. Each participant signs a written release, and the buyer holds a perpetual, irrevocable, worldwide licence, non-exclusive by default. Formats are RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd, with a median of 48 hours from signed spec to first validated batch where coverage is deep and three to four weeks where it is thin.

Both sources point to a habit this site recommends: decide which coverage is missing before choosing an hour count. See /guides/robot-pretraining-data and /custom-collection/pretraining-data-collection.