HiPHI's humanoid-motion release proposes a different starting point from simply collecting more internet video: measured body and object states, deliberately varied across a semantic collection plan. Its reported 617.5 hours are substantial, but they include left-right mirroring of 308.7 hours of original capture. The distinction is central to evaluating both its scale and its claim to improve robot control.
The paper was submitted on August 17, 2026, revised September 8 and, according to its research record, accepted at CoRL 2026. Creator-reported benchmarks connect data quality to learned tracking policies and selected Unitree G1 demonstrations. They do not show a general-purpose robot independently understanding and completing unfamiliar tasks.
What the release actually contains
Researchers used optical capture at 90 hertz with 132 performers and report sub-millimeter marker tracking. That precision describes the capture system, not a guarantee that every retargeted robot joint has sub-millimeter error. The release contains about 200.1 million frames, standardized 55-joint BVH body records, natural-language descriptions and semantic identifiers.
Released duration comprises 371.8 hours of body-only motion and 245.7 hours of human-object interaction, the latter 39.8% of the total. Synchronized object position and orientation tracks accompany high-resolution meshes, with one object entry per body frame. There are 40 canonical real objects across twelve categories, weighing 0.45–6.25 kilograms, plus mirrored mesh variants.
That structure preserves information a body-only clip loses. A sitting motion depends on a chair's surface; pushing a load changes balance and foot placement. Video can depict those events while requiring uncertain estimates of depth, occlusion and contact. Motion capture measures trajectories more directly, but a studio cannot match internet video's environments and spontaneity.
HiPHI therefore strengthens one layer of training data rather than replacing every source. Human demonstrations supply behavior, measured geometry constrains imitation, and robot-specific training must still translate the references into executable control.
FrameNet organizes what to capture, not just what to call it
The collection uses Berkeley FrameNet's event frames and lexical meanings as seeds. A Frame-LU pair distinguishes a word's physical meaning from another sense—for example running on foot versus running a company. HiPHI uses 22 frames and 214 motion units.
This matters because language labels and physical diversity are not equivalent. Wiping a forehead and shading the eyes can share a trajectory; “walking” can hide changes in speed, direction, stride, turning and height. Collection instructions vary measurable dimensions such as amplitude, rhythm, posture, contact position, load and object path.
The fifty most frequent Frame-LU labels account for 53.7% of duration, leaving 46.3% outside that group. The median unit has 24 performers; 154 units include at least ten. That gives both common movements and a deliberately sampled tail, while reducing dependence on one performer's style.
Coverage testing normalizes datasets to 23 keypoints at 30 frames per second, uses a shared unsupervised encoder and compares balanced samples. HiPHI occupies 1,620 projected cells versus BONES-SEED's 1,438, with effective occupancy of 1,443 versus 1,114. Effective occupancy reflects distribution rather than only whether a cell has one example.
The long-tail measure is 14.1% versus 10.7%. The researchers report a positive margin across tested seeds, projections and grid settings. These are results under that representation, not proof of universal coverage of the motion manifold. Balanced comparisons are more informative than hours alone, but a low-dimensional projection cannot measure every behavior.
Better contact geometry is not a force measurement
HiPHI reports 98.1% non-conflict and 95.7% near-surface grounding. The first tests whether the sampled skeleton avoids penetrating the object; the second tests proximity compatible with interaction. HIMO's corresponding figures are 97.6% and 79.0%, OMOMO's 90.6% and 50.4%.
A hand near a mesh need not exert the right force or maintain a stable grasp. These tests do not establish friction, tactile response or realistic loading. The interaction comparison also uses different available volumes: HiPHI's 245.7 released hours versus 21.6 evaluated object-track hours for HIMO and 9.8 for OMOMO.
Among datasets with comparable floor conventions, the creators report the lowest values on five selected smoothness/contact-artifact measures. HiPHI's 95th-percentile ground penetration is eight millimeters, versus eighteen for BONES-SEED and 111 for AMASS. Unsupported floating occupies 0.015% of evaluated duration; support-point drift is 64 millimeters per second. Dataset-level measurements are not promises about every clip.
Downstream interaction results are mixed. For pushing, HiPHI's body mean per-joint position error is 99.20 millimeters, worse than OMOMO's 66.09, while object position and orientation errors are better. In carrying, its body error is lower than both comparison results but object-position error is higher. Preserving an object's motion and matching a human pose can compete, so the evidence does not support winning every task.
Shared training makes the control test more informative
The team retargeted human references to Unitree G1, whose proportions and actuator limits differ from a person. Body-only policies used a DeepMimic-style reinforcement-learning pipeline. AMASS, LAFAN1, Motion-X++ and BONES-SEED underwent the same retargeting and training under matched data budgets.
HiPHI reportedly achieved the highest success and fastest convergence in three-hour and twenty-hour settings, with five independent training runs. A separate scaling experiment used unmirrored data from three to 300 hours, evaluated on held-out motion from four other datasets. Cross-dataset joint error declined as data grew; ten repeats supplied mean curves and variance bands.
Those experiments connect collection design to a control metric rather than only visual reconstruction. Their interpretation still depends on the shared retargeter, policy and evaluation choices. Different embodiments or training architectures could change the relative advantage.
Physical G1 trials included running, sitting, crawling, carrying a box, flipping and pulling a suitcase. Real actuators and sensing provide stronger executability evidence than simulation alone. They remain selected behaviors in controlled conditions, not demonstrations of autonomous object recognition, destination choice, planning or recovery in an unfamiliar workplace.
A tracking policy follows a reference while maintaining stability. A complete robot agent must also decide which reference is appropriate. Motions require embodiment-specific retargeting and safety checks; success on G1 does not transfer unchanged to every humanoid.
The release has explicit boundaries for adoption
All original capture is studio-based and single-person. It lacks the clutter, uneven ground, lighting change and surprises of natural environments, and excludes direct human-human interaction. Handovers or cooperative carrying would need information about multiple bodies and their changing relationships.
The release records kinematics, not contact forces or tactile signals. Similar visible trajectories can involve different resistance or grip pressure. That matters for fragile, slippery or deformable objects, where accurate motion alone is insufficient.
Mirroring expands useful left-right coverage but adds no independent performers or capture sessions. CoRL acceptance supplies peer review, while broader trust still requires outside reproduction of the creators' comparisons and hardware results.
The dataset uses CC BY-NC 4.0, so it should not be presented as unrestricted commercial training material. The public package excludes captured RGB video, audio, face images and voice, using numeric performer identifiers and generalized attributes. Motion itself can still contain individual patterns; absence of a face is not proof of zero identifiability.
HiPHI's defensible contribution is structured motion infrastructure: a larger set of precise, semantically organized references with synchronized objects and evidence of selected executable skills. Independent retraining, transfer beyond G1 and complementary force, touch or visual context would test how reusable that foundation becomes. The release advances the comparison between video breadth and measured physical structure without claiming to solve the rest of humanoid autonomy.