At a Glance
AWIGN specializes in capturing first-person point-of-view (POV) video and sensor data through wearable devices — footage designed specifically for training robotics AI systems. Their model relies on producing large volumes of richly annotated data that captures how humans naturally perform physical tasks, giving robotic learning systems the ground-truth they need to replicate real-world action sequences.
But building high-quality training data for robotics is not a volume problem alone. It is a precision problem. Every video must be annotated with structured detail — capturing not just what is happening, but the sequence of tasks, sub-actions, and contextual scenes that surround them. Doing this at scale, manually, quickly becomes the bottleneck.
AWIGN partnered with ThirdAI Automation to build an annotation pipeline that combines AI automation with human review — designed around the exact way their annotators worked and structured to scale as their business grew.
Key takeaways
03- 01AI pre-labels every uploaded video for scenes, tasks, and actions, so annotators review and refine instead of starting from a blank timeline.
- 02A purpose-built Scene → Task → Action taxonomy mirrors how first-person activity data is actually organised — not a generic bounding-box tool.
- 03Humans stay firmly in the loop: every published annotation is verified by a qualified reviewer before it is handed downstream.
The Challenge
For AWIGN, the challenge was not video capture — that part was handled reliably by their wearable devices. The challenge was what came next: transforming raw first-person footage into structured, high-fidelity training data that a robotics AI could effectively learn from.
Manual annotation was the default approach. Each video required an annotator to sit through the footage frame by frame, identifying scenes, tasks, and individual actions on a timeline — often multiple layers of annotation per single video. The output was precise, but the process was slow. As customer demand for training datasets grew, the annotation pipeline itself became the constraint. More footage meant more annotators, longer turnaround times, and rising costs — with no clear path to scaling without sacrificing quality.
Off-the-shelf annotation tools were built for general-purpose labelling — bounding boxes, object detection, and image classification. Very few were built for the specific rhythm of first-person action data: nested hierarchies of scene, task, and action; long-form video with continuous activity; frame-precise timeline annotations. AWIGN needed a platform designed around how their annotators actually worked.
A Purpose-Built Annotation Pipeline
Rather than adapting a generic tool, ThirdAI Automation built a video annotation platform tailored to AWIGN's exact workflow — combining AI-driven pre-labelling with structured human review.
At the heart of the system is a human-in-the-loop model. AI automatically pre-labels each uploaded video, identifying scene changes, tasks, and actions based on prompt-driven natural language instructions. Human annotators then review, correct, and refine those outputs, ensuring the final dataset meets the accuracy standards that robotics training requires.
Platform capabilities
- Structured annotation hierarchy — a purpose-built Scene → Task → Action taxonomy that mirrors how first-person activity data is naturally organised.
- AI pre-labelling with tiered accuracy — three sampling modes (Standard, High, Ultra) that let teams balance credit consumption against annotation granularity based on footage complexity.
- Prompt library — reusable natural-language annotation instructions, so the same guidelines can be applied consistently across projects and annotators.
- Team workflow with role-based access — dedicated Labeller, Reviewer, and Admin roles that allow projects to scale from a handful of annotators to distributed teams.
- Quality tracking built-in — labeller quality scores update automatically based on reviewer decisions, giving project managers a live view of dataset reliability.
Delivering Speed Without Sacrificing Precision
For a use case as specialised as robotics AI training, the temptation is often to trade accuracy for speed. The ThirdAI platform was built to eliminate that trade-off.
By automating the first pass of annotation, AI removed the mechanical work of identifying obvious scene boundaries and action sequences. By keeping humans firmly in the loop for review, the platform ensured that every published annotation was verified by a qualified reviewer before being handed downstream. The result was a pipeline that could handle rising volumes without diluting quality — the same rigour that manual annotation had always delivered, at a pace that matched growing customer demand.
The same rigour manual annotation had always delivered — at a pace that matched growing customer demand.
Beyond First-Person Video
While ThirdAI's engagement with AWIGN centred on first-person POV footage for robotics training, the underlying architecture applies to a broader range of use cases where structured video annotation matters:
- Wearable-captured activity datasets — for hand-tracking, gesture recognition, and human-motion research.
- Field-operations video training data — where nested action hierarchies (task → sub-task → step) reflect real workflow structure.
- Specialised annotation workflows — any use case where general-purpose tools force teams into rigid taxonomies rather than adapting to how the data is actually organised.
A Partnership Built on Iteration
Building a platform this specialized required working closely with AWIGN's team throughout the engagement. Every workflow decision — from annotation of timeline behaviour to the handling of overlapping actions — was shaped by continuous feedback from the annotators using the platform. Rather than being delivered as a fixed specification, the platform was continuously iterated, refined, and tuned to reflect how users worked.
AWIGN got a platform purpose-built to their use case — one that fit the way their annotators worked, matched the structure of their training data, and scaled with their business without requiring them to rebuild their pipeline from scratch.
The Outcome
For AWIGN, the outcome was a platform purpose-built to their use case — one that fit the way their annotators worked, matched the structure of their training data, and scaled with their business without requiring them to rebuild their pipeline from scratch.
For ThirdAI Automation, the engagement demonstrated something more foundational: that the right combination of AI automation and human review can unlock annotation velocity for even the most specialised data domains — from robotics training data to any use case where structured video annotation is the bottleneck.
