Robot Video Annotation Jobs

Embodied-AI data annotation is the work of labeling the raw sensor data that physical robots and self-driving systems learn to see the world through. Where a chatbot learns from labeled text, a warehouse robot, a humanoid, or an autonomous vehicle learns from labeled 3D point clouds, fused camera and LiDAR frames, and video of human hands and machines doing real tasks. Someone has to draw the cuboid around every pedestrian in a spinning LiDAR scan, color every pixel of drivable road, and tag the exact frame where a gripper closes on a cup. That someone is an embodied-data annotator, and it is still the most accessible remote entry point into robot-training work.

What the work actually is

Annotation for embodied AI is more three-dimensional and more physical than the text-and-image labeling that trains language models. The core task families:

• 3D bounding boxes (cuboids) in LiDAR point clouds. You place a tight 3D box around each vehicle, pedestrian, cyclist, and obstacle in a point cloud, then track that same object across a sequence of frames so the model learns motion. This is the backbone of autonomous-vehicle perception.
• Sensor fusion. The same scene is captured by LiDAR, radar, and several cameras at once. You label objects consistently across all of them so the model can fuse the views. This is harder, and pays more, than single-sensor work.
• Semantic and instance segmentation. Coloring in every point or pixel that belongs to a class (road, sidewalk, pole, person). Slower and more precise than boxes.
• Action, affordance, and trajectory labels on video. On robot-arm and egocentric video, you mark where a task begins and ends, which object is grasped, the path a hand or gripper follows, and whether an attempt succeeded. This is the labeling that feeds the vision-language-action models the humanoid labs are racing to build.

Embodied annotation rewards spatial reasoning and patience with 3D tools far more than writing skill. In the robot-training pipeline it sits right next to demonstration and data collection: they capture the data, you turn it into ground truth.

What it pays

Read from today's open roles. 17 open roles on RobotTraining.jobs are in embodied-AI annotation, and 7 of them state an hourly rate. Those rates typically run $7 to $75 an hour, with a median of $30 an hour. They come from individual postings with different duties and locations, not from a category-wide wage survey; the pay explorer and pay report show the full measured sample and its coverage.

An advertised hourly amount is gross compensation, not verified take-home earnings. Salary, equity, benefits, overtime and contractor expenses must be compared separately. We do not infer what a beginner or an expert earns by splitting a pay distribution into tiers.

For accepted-footage or task payments, include setup, recording, upload, review and any rework in your time calculation. An hour of accepted video is a unit of output, not necessarily an hour of labor. There is no guaranteed volume of accepted tasks or paid hours.

Who hires for it

Two very different channels hire embodied annotators.

The robot and AV builders themselves. They increasingly keep annotation in-house because label quality is a competitive edge. Figure AI lists a Data Quality Analyst alongside its robot operators; Skild AI hires master annotators; Tesla staffs data-operations roles for Optimus, and Agility Robotics and Boston Dynamics run their own data-ops teams. These pay US wages and sometimes carry benefits, but usually expect on-site or hybrid work in a specific city.

The annotation networks. These recruit remote contributors worldwide and price by geography. Scale AI, Sama, and Objectways run point-cloud and sensor-fusion pipelines behind many autonomous-vehicle and robotics programs. This is the accessible, work-from-anywhere door, but it is also where the geography gap and the labor-practice scrutiny concentrate: Scale AI was investigated by the US Department of Labor over how it classifies and pays its contractors; the department dropped that probe in May 2025, and private lawsuits over wages and classification were still pending when it did.

Tools, skills, and how to start

You do not need a degree. You need reliable internet, a computer that can smoothly render 3D point clouds (a discrete GPU helps for LiDAR work), and the patience to hit both a speed target and an accuracy target at the same time. The everyday tools are browser-based 3D suites: point-cloud cuboid editors, segmentation brushes, and video timelines for action labeling. Most networks gate you behind a short unpaid qualification task, then start you on simple 2D boxes before promoting you to 3D and sensor-fusion work, which is where the pay actually climbs.

A realistic path: pass a network's assessment, build speed on 2D and simple cuboid work, then qualify for sensor-fusion or QA and evaluation tiers, or apply directly to a robot lab's in-house data-operations team once you can show you understand 3D geometry and consistent, sequence-stable labeling. Many workers keep their hours full by pairing annotation with egocentric capture, which uses an overlapping skill set.

The trend that changes the floor

Auto-labeling is real, and it is compressing the simplest work. Models now pre-label a point cloud or pre-segment an image, and the human job shifts from drawing every box to reviewing, correcting, and resolving the edge cases the model gets wrong. That is good for pay per hour of skilled review and bad for anyone hoping to earn on sheer volume of easy boxes. The durable niches are 3D sensor fusion, ambiguous safety-critical scenes, and action or trajectory labeling on novel robot tasks: exactly the work that is hardest to automate. Before you commit your hours, compare the pay across every robot-training role on RobotTraining.jobs's live pay explorer.