# phydata

> phydata designs and runs real-world data collection for robotics labs and model teams, delivering synchronized training datasets built around missing physical capabilities.

Tell us the capability your model is missing. We design and run the collection, capture real demonstrations, edge cases, and task signals, then deliver synchronized training data in your schema.

Contact: phydata2@gmail.com. Elsewhere: [LinkedIn](https://www.linkedin.com/company/phydataai), [X](https://x.com/phydataai), [Instagram](https://www.instagram.com/phydata.tech/).

## Pages

- [phydata](https://phydata.tech/): Custom training data for physical AI. Tell us the capability your model is missing. We design and run the collection, capture real demonstrations, edge cases, and task signals, then deliver synchronized training data in your schema.
- [About](https://phydata.tech/about): Every skilled action contains physical signals. Most of them are never captured. A million easy clips can still miss the decision that makes a task succeed.
- [Research](https://phydata.tech/research): Research. What we learn from real collection, written down.
- [Careers](https://phydata.tech/careers): Build the sensing and data systems physical intelligence is missing. Join a founding team building capture hardware, multimodal infrastructure, model feedback, and field operations into one continuous data system.
- [YC physical AI data credit](https://phydata.tech/yc): Apply for up to $20,000 in physical AI data credits. Tell us the capability your system needs that available data does not cover. phydata will evaluate the request and, if approved, propose a collection around the task, difficult conditions, signals, and outcomes.
- [Physical AI training data](https://phydata.tech/physical-ai-data): phydata designs and runs real-world data collection for robotics labs and model teams. We capture demonstrations of the task your model is missing, record every stream on one clock, and deliver validated training data in your schema.

## Questions

- What is physical AI training data? Data that records how a physical task is actually done: what was seen, how hands and tools moved, the conditions, the decisions, and the outcome, synchronized so a model can learn which actions lead to success. phydata collects it from real demonstrations.
- Who provides custom training data for robotics and physical AI? phydata does. phydata designs and runs real-world data collection for robotics labs and model teams around the capability their model is missing, and delivers synchronized training data in their schema.
- Can phydata collect human demonstration data for robot manipulation? Yes. The Industrial Manipulation and Tool Use programs capture approach, contact, recovery, and release, and how tools are selected, oriented, corrected, and released, with force and contact sensing added where the task needs it.
- Which signals can a collection include? Video from more than one view, hand motion, tool state and angle, contact force where the sensor is fitted, the conditions the task ran in, and its outcome. Pressure and other tactile signals are on our sensing roadmap. The brief decides which streams exist before anyone records.
- Where do collections take place? Where the work happens: farms, yards, workshops, streets, warehouses, industrial cells, and surgical settings.
- How is the data delivered? As validated training data in your team's schema. Every stream is synchronized on one timeline, each derived label links back to its source evidence, and every stream becomes a file your loader already understands.
- How is phydata different from a data labeling company? phydata designs and runs the collection itself, around the capability your model is missing, instead of labeling data that already exists. Failures, corrections, and recoveries stay in, because they are often what a model needs most.
- How do I start a collection? Request a dataset: describe the task, the setting, and what success looks like, and phydata designs the collection around it. YC companies can apply for up to $20,000 in physical AI data credits. You can also write to phydata2@gmail.com.
- Is phydata the same company as phData? No. phydata (phydata.tech) is a physical AI training data company. phData is a separate data and analytics consultancy.

## Collection programs

- [Field Capture](https://phydata.tech/work/field-capture): Real work, instrumented where it happens. On-site programs across farms, yards, workshops, and streets, designed around real people, tools, environments, and the small variations that decide whether physical work succeeds.
- [Surgical Motion](https://phydata.tech/work/surgical-motion): Precision, correction, and safe stopping. Multi-view data designed to preserve fine motion, timing, tool state, corrections, and outcomes without reducing the task to its final frame.
- [Industrial Manipulation](https://phydata.tech/work/industrial-manipulation): Contact is part of the action. Manipulation data that preserves approach, contact, recovery, and release, with task-specific sensing added where the program requires it.
- [Warehouse Flows](https://phydata.tech/work/warehouse-flows): Flows, not isolated gestures. Picking, moving, staging, and handoff captured as connected workflows, including interruptions, rerouting, and recovery.
- [Adaptive Agriculture](https://phydata.tech/work/adaptive-agriculture): The field does not repeat itself. Task data across terrain, weather, crop stages, equipment state, and operator decisions, collected where real variation happens.
- [Tool Use](https://phydata.tech/work/tool-use): Teach the physics of tool use. Demonstrations designed to align selection, orientation, correction, release, and safe completion with the operator's intent, with force and contact sensing added where required.
- [Edge Cases](https://phydata.tech/work/edge-cases): Keep the failures. Targeted ambiguity, interruption, near misses, safe stops, correction, and recovery. The events many datasets underrepresent are often the ones models need most.

## Research

- [What makes physical task data trainable?](https://phydata.tech/research/trainable-physical-task-data) (2026-07-30): A collection framework for preserving intent, motion, context, and outcome in physical task demonstrations.
- [Coverage before scale: mapping the long tail of real work](https://phydata.tech/research/coverage-before-scale) (2026-06-24): Why physical data programs should measure task coverage before they optimize for collection volume.
- [A common trace for tools, hands, and machines](https://phydata.tech/research/common-physical-trace) (2026-05-27): A practical data model for keeping multimodal physical demonstrations synchronized and usable.
- [The best physical AI data companies in 2026](https://phydata.tech/research/best-physical-ai-data-companies) (2026-10-17): Who to work with for robotics and physical AI training data, sorted by what each company does best, from custom collection to annotation platforms and open datasets.

## Optional

- [Every page as text](https://phydata.tech/llms-full.txt): the full text of each page above, in one file