Short answer
Sanding, polishing and deburring depend on contact force, tool pressure and surface response, none of which ordinary video records. A useful dataset pairs synchronized video with force or tool-pressure signals and surface-condition labels. Existing footage can show technique and sequence; force-bearing data usually has to be commissioned, from instrumented tools or purpose-built rigs at manufacturers and finishing shops.
The buyer's problem
Finishing looks simple on video: a hand moves a tool across a surface. The skill is in what the video does not show. How hard the operator presses, how pressure changes at an edge, when the abrasive is spent, and how the surface looks and feels when the pass is done all determine the result. A model trained on footage alone can reproduce the motion pattern of a hand and miss the contact behavior that makes the finish acceptable.
This page helps a buyer decide what to source for finishing tasks and where video stops being enough. The wider question of recorded motion versus contact is covered in what human demonstration video can and cannot do for robot learning.
Task anatomy
Finishing covers several related operations with different data needs:
- Sanding: abrasive on flat or contoured surfaces, by hand, block or powered sander, with a grit progression.
- Polishing and buffing: compound applied to a wheel or pad, with dwell time and pressure controlling the result.
- Deburring: removing sharp edges or burrs from cut, cast or machined parts using files, scrapers or rotary tools.
- Surface preparation: cleaning, wiping and masking before coating.
Common steps are inspecting the surface, selecting abrasive or compound, fixturing or holding the part, working the surface in passes, checking the result by sight and touch, and repeating or moving on. Failure modes include over-sanding through a coating, burn marks from dwell, uneven material removal, missed edges, and dropped parts.
For the related manufacturing finishing step on molded and cast parts, see part deflashing and trimming.
Why force matters here
RH20T assembled more than 110,000 contact-rich real-world manipulation sequences, with visual, force, audio and action data for each sequence, and paired each robot sequence with a human demonstration video and language description (Fang et al., 2023). The relevant point for a finishing buyer is the design choice: a dataset built for contact-rich skills records force alongside video because video alone leaves the contact state unobserved.
Finishing is a clear case. Two operators can move a sander along the same path with different pressure and get different surfaces. Without a force or pressure signal, those two demonstrations look the same. See force-torque data for what the signal is and how it is recorded, and tactile sensing data for contact sensing at the fingertip.
What data helps each part
| Aspect | Data that helps | Typical source |
|---|---|---|
| Task sequence and technique | Fixed and wrist-camera video, step labels | Existing footage or commissioned recording |
| Contact pressure | Force-torque at the tool or part, or tool-side pressure logs | Instrumented tool or rig; commissioned |
| Tool state | Speed, current draw or pad condition from powered tools | Tool controller logs, if available |
| Surface outcome | Before and after images under consistent lighting; roughness or gloss readings | Quality records, commissioned imaging |
| Acceptance | Inspector verdicts, rework flags | Existing quality records |
Existing assets are strongest on the first and last rows. A finishing shop may hold training video, before-and-after photos and inspection results. They are rarely synchronized to any force signal, so what they can support is task understanding and outcome labeling.
Commissioned capture for finishing
Contact-rich demonstrations are normally designed rather than found. A capture plan would choose the parts and finishes, fit sensors to the tool or the part, record from chosen viewpoints and attach per-pass labels. Options include instrumenting a hand-held tool, mounting the workpiece on a force sensor, or using a handheld gripper or robot rig; see handheld gripper demonstrations and teleoperation data collection requirements.
Instrumenting a tool changes how it feels to use. Ask whether operators had time to adapt, and whether the added mass or cable altered technique. Sensor placement also matters: a force sensor under the workpiece measures a different thing from one between the hand and the tool.
Production shops carry their own constraints. Dust, coolant and compounds foul cameras and sensors, and machine-safety rules limit where equipment can be placed; see machine safety constraints for recording.
Illustrative finishing demonstration package
# Illustrative only - synthetic values, not an available dataset
package: surface_finishing_v0
site: site_021 (synthetic)
operation: orbital_sanding, grit sequence stated per pass
streams:
video: [fixed_oblique, tool_close_up, wrist_or_head_mounted], 60 fps
force_torque: tool-side or workpiece-side (stated per episode), 500 Hz, own clock
tool_log: speed and current if the tool exposes them
sync: shared clock or recorded sync event; max drift declared per episode
calibration: sensor frame, tool center point and camera extrinsics stored
labels:
pass_segments: start/end per pass, grit, direction
surface_outcome: pre/post images, inspector verdict (accept / rework)
metadata: part_id part_B31, material, operator_id operator_A12
excluded: robot joint states (unless a robot or handheld rig was used)
rights: training and evaluation use; operator consent recorded
Limitations and additional preparation
- Embodiment. A human hand tolerates and senses contact differently from a robot end effector. Force traces describe the human's technique, so they need retargeting or a robot rig to become action data.
- Subjective outcomes. Whether a surface is acceptable can depend on an inspector's judgment, which should be recorded as a label with the criteria behind it.
- Tool and abrasive variation. Wear changes behavior across a session. Record abrasive condition or replacement events.
- Synchronization. Force at hundreds of hertz and video at tens of frames per second need a shared time base; see sensor synchronization requirements and calibration requirements.
- Small samples of materials. A dataset on one alloy and finish will not tell you how a robot behaves on another.
Permissions and provenance
Parts and finishes may belong to customers, and drawings or surface specifications may be restricted; see manufacturing data confidentiality. Operator video and force traces both describe a person's technique, so consent should address both. This is not legal advice.
Choosing a capture route
Decide the route by the question you need answered. If you only need to know how experienced finishers sequence passes and judge a surface, existing or lightly instrumented video with step labels may be enough. If you need to know how much pressure produces an acceptable result, you need force or tool-pressure data synchronized to video. If you intend to train a robot policy, plan for demonstrations recorded in the robot's action space through teleoperation or a handheld rig, and treat human finishing data as supporting context. Whichever route you choose, state which surfaces, geometries and abrasives the data must cover, since these change the contact behavior more than the operator does.
Questions to ask suppliers
- Which finishing operations, materials and grit or compound schedules are represented?
- Is force or pressure recorded, where is the sensor placed, and at what sampling rate?
- How are force, tool logs and video aligned, and what drift is declared?
- Were tools instrumented specifically, and how did operators report the effect on technique?
- What defines an accepted surface, and who judged each outcome?
- Are before-and-after images captured under consistent lighting?
- How many operators and parts appear, and how does technique vary between operators?
- Can the shop provide coverage of rework passes and mistakes?
- Which part geometries or customer specifications must be excluded?
Discuss your requirements
If your finishing tasks depend on contact data, describe the operations, sensing and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. Where no existing asset fits, SourceX can scope a commissioned collection. A request does not guarantee a match. See the humanoid training data guide and the manufacturing hub for related tasks.
Discuss your humanoid training-data requirements with SourceX
Sources
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — RH20T: 110k+ contact-rich sequences with visual, force, audio and action data, paired human video.