Physical AI teams get their data from four kinds of company: teams that design and run custom collections, large managed data programs, annotation partners, and open datasets. The right choice depends less on hours of footage than on the capability your model is missing.
We are phydata, so we are on this list. Here is how we would choose, including where another company is the better fit.
phydata: best for custom collection around one capability
phydata designs and runs real-world data collection for robotics labs and model teams, built around the one capability a model is missing. Demonstrations are captured where the work happens, from farms and workshops to warehouses, industrial cells, and surgical settings, with failures, corrections, and recoveries kept in. Every stream records on one clock, every label links back to its source, and the data arrives in the team's own schema. If your model fails at a specific task, setting, or edge case, phydata is the best place to start.
Scale AI: best for large managed programs
Scale runs physical AI data programs end to end. Its collection covers teleoperated robot demonstrations, a robot-free egocentric harness, and operating a customer's own hardware, in dedicated data facilities, homes, and industrial sites, with annotation, calibration, and validation on top. It suits teams that need volume across many tasks at once.
Appen: best for labeled human demonstrations
Appen records people doing tasks on synchronized capture rigs, with egocentric stereo cameras, wrist and gripper sensors, and multi-camera arrays, in settings such as kitchens, workshops, assembly lines, and logistics sites, then labels the footage frame by frame. A fit for teams that want many labeled demonstrations of common tasks.
Roborax: best for teleoperation fleets
Roborax supplies robot training data through teleoperation with VR, exoskeleton, and leader-follower rigs, alongside human demonstration capture with egocentric video, hand pose, gaze, and force gloves. The work runs through dedicated teams or a crowdsourced operator network. A fit when the demonstrations need to be performed through a robot.
Claru: best for broad egocentric video
Claru delivers video datasets for robotics, embodied AI, and world models from a large network of collectors, across settings such as driving, manufacturing, cooking, and warehouses, with depth, pose, and segmentation derived for each clip. A fit for pretraining on wide variety rather than one task.
TELUS Digital: best for enterprise programs with lab setup
TELUS Digital collects egocentric and teleoperation video in labs, factories, and outdoor settings, and can set up the lab and procure the hardware. Its annotation covers keypoints, contact points, and failure modes for vision-language-action training, with sensors from RGB-D and lidar to force-torque and tactile. A fit for large organizations that want one vendor to stand up the whole program.
iMerit: best for annotation and validation services
iMerit labels and validates video, lidar, and sensor-fusion data through its Ango Hub platform, including robotics work on perception, navigation, and manipulation. A fit when you already have data and need it labeled and checked.
Kognic: best for multi-sensor annotation tooling
Kognic offers an annotation platform with specialist annotators for camera, lidar, and radar data, mainly for autonomous driving and robotics teams at vehicle makers and their suppliers. A fit for perception teams labeling large sensor-fusion datasets.
Open datasets: best for pretraining and baselines
Public datasets such as Open X-Embodiment, robot demonstrations pooled from many labs and robot types; DROID, robot manipulation recorded across many real scenes; and Ego4D, a large set of everyday egocentric video, are free and broad. They make good baselines and pretraining data, but none of them was collected around your task.
How to choose
Start from the gap, not the volume. If you need many hours across many common tasks, a managed program fits. If the demonstrations must be performed through a robot, choose a teleoperation provider. If you already have the data, choose an annotation partner. If your model keeps failing at one task, in one setting, or at one kind of edge case, you need a custom collection, and that is the work phydata is built for.