Choosing a robot for data collection: buy for uptime, not for precision
Repeatability and payload matter less than duty cycle, repair time, and readable joint state. How to pick arms, grippers, and cameras for a collection rig.
A data collection robot is a production machine, but the specs that win industrial sales are not the specs that decide how much data you get. Repeatability in fractions of a millimeter, stiffness, and cycle time answer a manufacturing question. Collection asks a different one: how many usable hours of episodes can this rig produce per week, and how long does it sit broken when something goes wrong.
The three specifications that decide throughput
Everything else is secondary to these.
- Continuous duty behavior. A collection rig runs for hours at a high control rate, and an arm rated for intermittent industrial cycles can derate or overheat under that pattern. Ask for the continuous rating, not the peak.
- Time to repair. Crash damage is normal, especially with operators who are still learning. The question is whether a replacement wrist, finger, or cable is a stocked part you swap in an afternoon, or a factory return that costs weeks of collection time.
- State and command access. You need joint positions, velocities, and ideally commanded torque at the control rate, plus gripper state on the same clock. An arm that only exposes end-effector poses over a slow interface limits what you can train on and makes every diagnosis harder.
Why precision is the wrong target
A repeatability figure of a few tenths of a millimeter describes the arm returning to the same point under the same conditions. Collection does not need that. It needs the arm to behave the same way across a shift and a state stream that is internally consistent from episode to episode. A cheaper six-axis arm with a documented, stable interface will usually out-collect an expensive seven-axis arm with a closed one.
Seven axes have one genuine advantage: the redundant joint makes it easier to avoid joint limits and keep a wrist camera oriented while reaching around an obstacle. If your tasks need that, the extra axis earns its cost. Torque-sensing joints matter if force data is part of your dataset, and they raise the price and slow repair.
The gripper decides which tasks are possible
A parallel-jaw gripper with position and force feedback covers a large share of pick-and-place work and reports state worth recording. A Robotiq 2F-85 class gripper, with roughly an 85 mm stroke, is the common reference point, and wider-stroke versions exist for larger objects. What matters for data is whether the gripper reports its width and whether it can tell you a grasp was detected. If it only accepts an open or close command and reports nothing back, you lose the most useful signal for labeling grasp events later.
Swapping grippers is a calibration event, so plan the swap at a session boundary and re-run the camera-to-arm calibration afterwards. For anything that requires adjusting an object inside the hand, a two-finger gripper cannot produce the data at all. That is a dexterous hand project with different tradeoffs, and it should be scoped separately.
Cameras: mount for the policy, not for the demo video
Third-person cameras show the arm and the whole workspace, which helps a policy learn spatial relations, but the manipulated object often occupies a small fraction of the image. Wrist-mounted cameras keep the object large and centered, at the cost of global context and occasional occlusion by the arm. A common configuration is two fixed cameras framing the workspace from different angles plus one wrist camera, and the wrist view is worth adding only if the task needs fine alignment the fixed views cannot resolve.
Depth cameras such as the RealSense D400 series or the ZED series give geometry cheaply, and the D455 and ZED 2i use global-shutter sensors for both streams, which matters when the arm moves fast. Record RGB and depth when the sensor provides both and decide at training time which to use. Mount cameras to the table or a rigid frame, never a tripod that gets nudged.
Buy a fleet instead of a hero
If the budget covers one high-end arm or three mid-range arms, the three-arm option usually produces more data and keeps producing when one station is down. Standardize ruthlessly: same arm model, same gripper, same cameras, same mounting plate, same control software at every station, because mixed hardware multiplies calibration work and quietly fragments the dataset into incompatible halves. Keep a spare gripper, spare camera cables, and spare fasteners on the shelf, since in a collection project a missing cable is a lost day.