Glossary
Short, accurate definitions of terms that come up when teams talk about training data for robotics and Physical AI, and where Zenyth's own data fits each one.
Egocentric data is video and sensor data recorded from a first-person point of view, typically a head-mounted or body-worn camera, rather than a fixed, third-person camera looking on.
Every session Zenyth captures is recorded from the worker's own point of view, with video and motion data synchronized to one timeline across the environments we work in.
Embodied AI describes an AI system built into a physical body, a robot or other agent that has to sense its surroundings and act on them, rather than a model that only processes text or images in isolation.
Zenyth uses the closely related term "Physical AI" for this same category of system; see that entry below.
Exocentric video is footage recorded from a third-person, external viewpoint, such as a fixed room or overhead camera, as opposed to a first-person "egocentric" viewpoint.
Zenyth's own capture is egocentric by default: recorded from the worker's own point of view rather than a fixed external camera.
Hand-object interaction data records how hands grasp, move and manipulate a specific object, often including grip type, contact points and the sequence of the interaction in detail.
Zenyth's egocentric captures record this natively: the camera travels with the person doing the task, so hands and the objects they're working with stay in frame throughout the environments we cover.
Human demonstration data is data recorded from a real person performing a task, used to teach a model what the task looks like and how it's done.
This is what Zenyth collects: real people doing real physical tasks in real working environments, not staged or simulated performances.
An IMU (inertial measurement unit) is a sensor that measures motion: acceleration, rotation and orientation, commonly worn or mounted to record how a body or object moves through space.
Zenyth pairs IMU-based motion data with its egocentric video so the visual record and the underlying movement are captured on one synchronized timeline.
Manipulation data records how an object is picked up, moved, adjusted or otherwise physically handled during a task, a core input robots need to learn to do the same.
This is a large share of what Zenyth captures: sorting, packing, cutting and positioning tasks across the environments we work in.
Multimodal training data combines more than one type of signal, such as video, motion and text, into a single synchronized dataset rather than treating each as separate.
This describes Zenyth's own output directly: synchronized video and motion data, structured with task and session metadata to a client's specification.
Physical AI refers to AI systems that perceive and act in the physical world: robots and other embodied agents that combine sensing (vision, touch, motion) with real-world action, as distinct from AI systems that only process text or images.
Physical AI models learn from real demonstrations, not just text and images; that reasoning is behind everything Zenyth collects.
A robot foundation model is a large model trained on broad demonstration data and then adapted to specific robots or tasks, similar in spirit to a foundation model in language or vision but trained to output physical actions.
VLA models (see VLA training data below) are one common type of robot foundation model.
The sim-to-real gap is the drop in a model's performance when it moves from a simulated training environment to the real world, caused by everything a simulation can't fully reproduce, such as lighting, material behavior, friction and human unpredictability.
Real-world captured data, rather than simulated data, is one direct way teams address this gap, since it reflects the physical world's actual variation instead of an approximation of it.
Teleoperation data is recorded from a robot being remotely controlled by a human operator, capturing the operator's control inputs and the robot's resulting motion.
Zenyth's own capture is different: it records a person directly performing a task with their own body, not a human operating a robot at a distance.
VLA (vision-language-action) models take in a visual observation and, usually, a language instruction, and output the action a robot should take next; VLA training data is the demonstration data used to train them.
Zenyth's core capture, synchronized egocentric video and motion data, covers the vision and action halves of that pairing; language-instruction annotation is scoped to each client's project through our existing annotation and structuring service, not delivered as a standing pre-packaged corpus.
Curious what this looks like in practice?
Get in touch
Tell us the task, environment, modalities and scale. We'll design the collection around your requirements.
Start a paid pilot