Flikforge launches two-tier EmbodiedAI data family for physical AI

Flikforge introduced EmbodiedAI, a two-tier training data family designed to give frontier labs and robotics teams rights-cleared, sensor-synchronized data for physical AI. The launch targets a market that has plenty of raw video but still lacks usable, commercially licensed training signal.

Why it matters: - Physical AI teams need curated training data that is usable in production, not just more video. - Flikforge is positioning EmbodiedAI as a way to cut the hidden cost of filtering, validating, labeling, and packaging egocentric footage into model-ready data. - The launch targets frontier labs, robotics teams, humanoid developers, autonomous systems, industrial automation, healthcare AI, defense, and world-model builders.

What happened: - Flikforge announced the EmbodiedAI family on July 9, 2026, in San Francisco. - The release introduces two tiers: EmbodiedAI-500k and EmbodiedAI-75k. - EmbodiedAI-500k is a curated corpus of more than 500,000 hours of rights-cleared human activity video. - EmbodiedAI-75k is a 75,000-hour engineered depth tier drawn from that corpus and paired with robot-native capture. - Flikforge says the family is built to move customers beyond raw video volume toward rights-cleared, model-ready training signal.

The details: - EmbodiedAI-500k is designed as a broad foundation layer for robot foundation models and world models. - The corpus is deduplicated, quality-filtered, coverage-scored, densely labeled, and delivered with customizable metadata. - Flikforge says the data includes full provenance and chain-of-title verification. - EmbodiedAI-75k delivers more than 3 million training-ready segments. - Those segments are organized around 1,000-plus robotic primitive skills. - Example skills include opening a screw-top jar, opening an appliance door, and picking a part from a bin. - The skill set spans Household, Commercial, Public Service, Industrial, and Entertainment & Companionship domains. - Flikforge says each skill is defined across variation axes including objects, surfaces, environments, lighting, handedness, skin tone, clutter, failure modes, recovery behavior, and task transitions. - The depth tier is designed to support composed long-horizon task sequences as robots move from single actions to full jobs. - The tier mirrors a training stack that combines coverage-scored human egocentric segments with synchronized sensor data and wrist-camera coverage, plus about 450,000 teleoperation segments with joint data, LiDAR, haptic force, and tactile feedback. - Flikforge says every segment passes through coverage and diversity scoring, sensor synchronization, normalization, dense sub-action indexing, and labeling. - Delivery formats include MCAP, LeRobot, and RLDS. - Buyers can license the full corpus, a single tier, a domain, or a custom skill list.

Between the lines: - The launch reflects a broader shift in robotics data strategy from more footage to better coverage. - Flikforge is leaning on research that links policy generalization to diversity of environments and objects rather than raw demonstration count. - The company is also betting that rights-cleared, provenance-rich data will matter more as commercial deployment increases. - Jeff Allen, founder and CEO of Flikforge, framed the market as a coverage problem rather than a volume problem.

What's next: - Qualified buyers can request the EmbodiedAI family and custom data schemas. - Flikforge is also directing readers to its companion Field Notes analysis, The Great Robot Data Glut, for more context on the egocentric video market shift. - The company is offering the data family to teams building physical AI systems that need both broad pretraining and skill-specific fine-tuning.

The bottom line: - Flikforge is trying to sell a new category of AI training data: curated, licensed, sensor-aligned signal built for robotics, not just piles of raw video.