Project Cursa - Robot Manipulation Video Annotator (V2)

The work

Label short robot manipulation videos to support robotics AI training. Produce precise, structured descriptions of observed actions rather than descriptions of what the robot was supposed to achieve. Strong written English, close observation, and consistent use of a detailed style guide are central to the work.

Responsibilities

• Watch each video through three synchronized views: one overhead camera and a camera mounted on each robot wrist.
• Divide videos into time segments and write clear natural-language descriptions for each segment.
• Apply atomic-motion, skill/subtask, and task/goal labels to applicable segments using the detail levels below.
• Cover every moment with labels at two or more levels, including idle time and pauses; leave no gaps.
• Describe exactly what occurs, including dropped objects, failed grasps, and slipped grips. Accuracy matters more than presenting successful outcomes.
• Cross-reference every camera view. Establish the scene and object identity overhead, then verify contact and grasp details in the wrist views.
• Apply the style guide consistently across episodes, including action vocabulary, spatial relationships, object descriptions, and manner of movement.
• Attend periodic calibration sessions to align labels with the team and client reference examples.

Labeling levels

• Atomic motion: a few seconds covering one small movement, such as closing fingers around a red handle.
• Skill/subtask: several seconds to approximately 20 seconds covering a meaningful action, such as picking up a red block by its edge.
• Task/goal: up to approximately one minute describing the purpose of a sequence, such as placing all blocks in a container.

Required skills

• Strong written English for dozens of short, precise descriptive sentences per video, with varied language rather than repeated phrasing.
• Sharp attention to small differences: a successful grasp versus a fumble, pushing versus dragging, or the exact object part touched.
• Comfort following a detailed, structured style guide consistently, including ambiguous and edge-case situations.
• Basic spatial and mechanical description skills, including left/right, above/below, and object parts such as handles, lids, and edges.
• Reliable, self-directed work habits, with focused independent work and periodic check-ins rather than close supervision.

Preferred background

• Video annotation, data labeling, transcription, or QA experience.
• Robotics terms such as grippers, end-effectors, and manipulation are helpful but not required; the style guide is self-contained.
• Experience with annotation tools such as Label Studio.

Arrangement and support

The work is fully remote and project-based, with flexibility over when and how much you contribute. Optional AI and Large Language Model workshops, responsive guidance, and a contributor community are available. No coding is required. Registration begins with a few introductory questions for the contributor database.