Robotics Data Release Tracker 2026: Human Video and Robot Trajectories

Robotics Data Is Scaling Along Two Different Axes

Source-backed context[1] DreamDojo official project and paper[2] AgiBotWorld-Beta official dataset card[3] Open X-Embodiment official project[4] DROID official dataset project

Human egocentric video now reaches tens of thousands of published hours, while robot-native collections reach one million or more trajectories. DreamDojo-HV reports 44,711 hours of human video. AgiBotWorld-Beta and Open X-Embodiment each report more than one million robot trajectories. DROID reports 76,000 trajectories across 350 hours.

Those top-line numbers cannot be ranked on one axis. Human video supplies visual, semantic, behavioral, and physical-world breadth. Robot demonstrations supply embodiment-specific states and actions. A trajectory is not a fixed amount of time, and an hour of human video is not a control sequence.

Human Video Provides Breadth Before Action Grounding

Source-backed context[1] DreamDojo official project and paper[5] Egocentric-10K official dataset card[6] Ego4D official project[7] EgoLive paper[8] EgoVerse official project[9] Ego-Exo4D V2 documentation[10] Apple EgoDex repository[11] HoloAssist official project[12] EPIC-KITCHENS official project

Human egocentric sources are strongest for world-model pretraining, task decomposition, language grounding, object-state understanding, hand-object interaction, visual representation learning, and motion priors. They normally stop short of robot-native joint state, gripper commands, rewards, or contact forces.

The practical bridge can be post-training on robot actions, a latent-action model, retargeted human motion, cross-embodiment co-training, or a separate policy-learning stage. The bridge must be stated; calling human video robot-action data is technically false.

Robot-Native Releases Add Control at a Different Cost

Source-backed context[2] AgiBotWorld-Beta official dataset card[3] Open X-Embodiment official project[4] DROID official dataset project

AgiBotWorld-Beta exposes action and proprioceptive structures at enormous scale, Open X-Embodiment standardizes data from 22 embodiments, and DROID holds hardware constant while increasing scene diversity. These sources are closer to policy learning because they carry a control or state signal.

They also inherit embodiment, teleoperation, schema, hardware, and component-license constraints. A million pooled trajectories do not guarantee compatibility with one target robot, and a single-platform dataset does not automatically generalize across embodiments.

Access and Rights Still Fragment the Market

Source-backed context[1] DreamDojo official project and paper[2] AgiBotWorld-Beta official dataset card[3] Open X-Embodiment official project[5] Egocentric-10K official dataset card[8] EgoVerse official project[11] HoloAssist official project[12] EPIC-KITCHENS official project

The tracked releases use signed agreements, gated dataset cards, non-commercial Creative Commons terms, component-specific licenses, public downloads, marketplaces, and project pages that do not establish complete corpus rights. Code and data remain separate legal objects.

A procurement record should pin the exact release, source agreement, commercial and model-training permissions, derivative rules, vendor access, redistribution, retention, and any obligation inherited from a pooled component dataset.

The Tracker Is Built to Change

Implementation guidanceEGXO guidance for translating the research into a project specification.

This is a dated release register, not a permanent ranking. EGXO reviews the CSV monthly and after material first-party announcements. New entries need an official source, preserved unit, access signal, license signal, and a limitation that prevents the headline figure from becoming empty hype.

• Use publisher scale as a discovery signal, not a usable-yield metric
• Compare hours with hours and trajectories with trajectories
• Separate human observation data from robot action data
• Verify the current license and access route before ingestion
• Pin the release and reproduce one representative loader test