Wrist camera
A wrist camera is a camera mounted on a robot's wrist or gripper, so it moves with the end-effector and gives a close-up view of the object being grasped. This setup is called eye-in-hand, as opposed to eye-to-hand cameras fixed in the scene. Many robot learning datasets, including DROID and LIBERO, pair a wrist camera with one or more fixed cameras.
In this article
Building visual or physical AI?
Let's talk.
What is a wrist camera?
A fixed camera sees the whole scene but often loses the object behind the arm at the moment that matters. A wrist camera sees the opposite: a close view of the object and the gripper, from the hand's point of view, as the robot reaches and grasps. That is why many manipulation datasets record both.
Robotics calls the two setups eye-in-hand, where the camera rides on the robot, and eye-to-hand, where the camera is fixed in the environment and watches the robot. DROID records one wrist camera and two exterior cameras per episode, and LIBERO pairs an agent-view camera with an eye-in-hand camera.
Key takeaways
- A wrist camera moves with the gripper and shows the object up close during a grasp.
- Fixed cameras show the whole scene, so most datasets record both views.
- Relating a wrist view to the robot's coordinates takes hand-eye calibration.
How it works
The camera is mounted on the last link of the arm or on the gripper itself, usually pointing along the gripper's approach direction. Because it moves, its pose changes every frame. Hand-eye calibration estimates the fixed transform between the camera and the end-effector, and combining that with the arm's forward kinematics gives the camera's pose in the world at every timestep. A learned policy can also use the wrist images directly, without explicit calibration, which many imitation learning setups do.
Watch the grasp from the wrist, not just the room
The wrist view shows what fixed cameras miss, and it has its own failure modes. Watch for:
- the object leaving the wrist view before the grasp closes
- the gripper hiding the object at the moment of grasp
- grasps that look clean from the agent camera and off-center from the wrist
In Voxel51, the LIBERO dataset holds 1,693 simulated episodes, each with an agent camera and a wrist camera slice.
ICRA 2026 Robotics LIBERO
1,693 episodes
Grouped samples from the LIBERO lifelong robot learning benchmark, with agent and wrist camera slices per episode.
Try this: Pick an episode and switch between the agent and wrist camera slices at the grasp. Note what only the wrist camera shows.
Explore LIBERO in Voxel51
Free with a work or .edu email.
Go deeper
Why it matters
The wrist view is often where grasp failures are visible. A fixed camera may show a clean pick, while the wrist camera shows the object slipping or the gripper closing off-center. Wrist frames bring their own problems too: motion blur during fast moves, the gripper or the object filling the frame, and timestamps that drift from the fixed cameras. Reviewing the wrist and fixed views together on one timeline, with sensor synchronization checked, is how those problems get caught.
Frequently asked questions
What is the difference between eye-in-hand and eye-to-hand?
Eye-in-hand means the camera is mounted on the robot and moves with the gripper. Eye-to-hand means the camera is fixed in the environment and watches the robot work.
Why do robot datasets record a wrist camera?
Because it shows the object and gripper up close at the moment of contact, which fixed cameras often miss. Many imitation learning policies use the wrist view alongside a fixed view for that reason.
Does a wrist camera need calibration?
For tasks that convert what the camera sees into robot coordinates, yes. Hand-eye calibration estimates the camera's pose relative to the gripper. Learned policies that map images straight to actions can work without it.
Related terms

Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 19, 2026