Introducing the closed-loop platform for Physical AI, and its first live capability.
Today we're introducing IntelliPhylo — the closed-loop platform for Physical AI.
The name joins intelligence with the Greek phylon — lineage. It carries our purpose: physical AI shouldn't be trained once; it should evolve. IntelliPhylo runs the loop that makes that happen — plan the data, assure its quality, train, evaluate on real tasks, mine the failures, and let what you learned script the next round of collection. Every cycle is a generation. Every generation, measurably better.
The loop's first capability is now open for access: DataProof — evidence-backed quality evaluation for embodied-AI data.
Here's the problem it removes. A batch of training data arrives — from a capture partner, a public benchmark, your own fleet — and days disappear into one question: can we trust it? Answer wrong, and you pay in GPU-hours and a regression you catch too late.
DataProof answers in minutes, with proof:
→ Four embodied data types — egocentric demonstration video, handheld gripper recordings, teleoperation data, and video Q&A — scored against a 122-point standard calibrated on real-world datasets.
→ It reads file contents, not filenames. A depth stream mislabeled "rgb" gets caught; nested archives get opened to the bottom. Real-world data arrives messy — DataProof understands it first, measures it second.
→ Terabyte-scale, evaluated in place. Nothing copied out of your storage; 100,000 files scanned in seconds.
→ Every score traceable to evidence — the exact files examined, per-scene breakdowns, model-type suitability guidance, and machine-readable instructions for what to collect next.
→ Scores, never verdicts. Your data is never "bad" — it scores X, here's why, and here's exactly what would raise it.
Two capabilities arrive in upcoming releases: AI-assisted scoring of judgment-based checks (with human review built in), and native readers for RLDS/rosbag ecosystem formats. Until then, they're labeled honestly rather than scored.
DataProof is the first gate of the loop — the rest of IntelliPhylo's modules follow it through Early Access. If deciding what your data is worth costs your team days, we'd like to give you those days back.