Short answer
Post-processing after 3D printing, such as removing parts from the build, taking off supports and finishing surfaces, is variable manual work done in service bureaus and print shops. Recordings of it can support task understanding, visual learning and evaluation. Learning to carry out the task on a humanoid needs robot or teleoperation data, and force-sensitive steps need instrumented capture.
Why this task is worth a separate look
Printing is automated; much of what comes after is not. A finished build often needs someone to remove parts from a plate or powder bed, clear loose material, snap or cut away support structures, and finish surfaces. Because each build can hold different geometry, the work resists the fixed-fixture logic of a conventional assembly line. That is both the appeal and the difficulty for a humanoid program: high variety, small batches, and many judgment calls about how much force a delicate feature will tolerate.
The processes involved differ by printing technology, and a buyer should not assume one workflow. Which steps apply at a given bureau is something to ask, not infer. The sections below describe the task anatomy generally and flag where a supplier's practice will vary.
Task anatomy
| Phase | What the worker typically does | Why it matters for data |
|---|---|---|
| Unload | Lifts a build plate or removes parts from a build volume | Large objects, two-handed carrying, awkward reach |
| Clear loose material | Brushes, blows or shakes off residue or unused material | Fine particles degrade video; depends on process |
| Detach from plate | Pries, cuts or releases parts | Force-sensitive; parts can be fragile |
| Remove supports | Snaps, clips, cuts or dissolves supports | Highly part-specific; hand and tool occluded |
| Finish surfaces | Sands, tumbles, sprays, dyes or smooths | Overlaps with sanding, polishing and finishing and deflashing and trimming |
| Inspect and sort | Checks features, groups parts by order | Needs labeling of pass/fail and destination |
Some phases involve powders, resins or solvents depending on the process. A buyer building a data plan should ask what protective equipment and handling controls the bureau uses, and look at how chemical hazard information is represented; see safety data sheets as handling constraints. That is a diligence question about the supplier's practice, not a statement about any regulation.
What recordings can support, and where capture is hard
Video of these steps may support step segmentation, recognition of tools and parts, language grounding of what was done and why, and evaluation of a model's task understanding. It can also show the variety that makes this work hard: the same support-removal step looks different on every geometry.
Tool use is a large part of the work; see hand tool use demonstrations for how tool handling is recorded. Capture is awkward for specific reasons. Supports are often small and sit inside the part's geometry, so hands and tools hide them from a head-mounted camera. Loose residue or airborne particles may foul lenses or drift across frames. Finishing steps involve repetitive strokes where the useful signal is pressure and angle, neither visible in video. Handling decisions, such as "this feature will snap, go slowly," are tacit and rarely narrated unless a protocol asks for it; narrated demonstrations can capture some of that reasoning.
Existing recordings versus commissioned capture
Service bureaus may hold training clips, quality photos of finished parts, or process-monitoring footage, though a buyer should ask rather than assume. Existing material was recorded for other purposes, so viewpoints and labels will rarely match a training need; see existing footage vs commissioned recording.
Commissioned capture lets a buyer choose viewpoints, tasks and labels, but it takes site access, demonstrators and consent, and needs a clear plan for what each take shows. A mixed route is often sensible: existing footage to understand variety, a smaller designed capture for the steps that matter. If recordings are made on a working floor, consult recording protocol on live work sites.
Illustrative capture package for support removal
This is an Illustrative package with synthetic values showing the fields a buyer could request for one step. Change it to suit your architecture and task.
# Illustrative only - not an available dataset
package: am_support_removal_v0
task: remove supports from printed parts, then place parts in a sort tray
sites: [bureau_A, bureau_B] # synthetic
views: [head_mounted, fixed_overhead_bench, close_up_bench_cam]
per_take_metadata:
part_family: [bracket, housing, lattice] # example categories
process_family: [powder_bed, resin, filament]
support_style: [tree, block, dissolvable]
tool_used: [hand_only, flush_cutter, pick, deburr_tool]
outcome: [clean, feature_damaged, retry_needed]
labels: [step_boundaries, tool_ids, part_ids, free_text_instruction]
not_included: [robot actions, joint states, contact forces]
rights: training and internal evaluation; supplier confirms authority over part designs shown
The part_family row matters because customer part designs can be confidential. A bureau may not be able to license footage that shows a client's geometry.
What this data cannot support
Video does not provide joint states, commanded actions or contact forces. Visuomotor policies are trained on demonstrations that pair observations with actions; one example is Diffusion Policy, which frames the policy as conditional denoising diffusion (Chi et al., 2023). That paper is a method reference only. It does not cover additive-manufacturing tasks, and nothing in it says what data post-processing needs. For a humanoid to perform support removal or finishing, a buyer should expect to need robot or teleoperation demonstrations, and for pressure-dependent steps, instrumented tools or force sensing.
Footage alone also cannot tell you whether a part was damaged beyond what is visible, whether a support was fully cleared, or what tolerance the customer required. Those labels have to come from inspection records or annotators.
Finally, one bureau's workflow may not transfer to another's. Technology, materials and part mix change the task. Collect enough variety, and evaluate on held-out sites and part families, as covered in rework and repair operations for related variability.
Permissions
Recordings may show workers, customer part designs and proprietary build settings on screens. Blurring or masking reduces visibility but does not settle whether the bureau may license the material or whether workers were given appropriate notice. Ask how customer confidentiality terms are handled. This is not legal advice.
Questions to ask suppliers
- Which printing processes do you run, and which post-processing steps are manual for each?
- Do you hold any recordings of unloading, support removal or finishing, and what were they made for?
- How varied is the part mix across a typical month, and can you describe it by family without disclosing designs?
- What protective equipment and handling controls apply to powders, resins or solvents at each step?
- Can customer geometry be excluded from frame, and at what cost to task visibility?
- Do operators narrate or label tricky steps, such as when a support will snap?
- Do you record outcomes such as feature damage or rework per part?
- Would you allow a designed recording session on a stopped or low-volume shift?
- Who can authorize licensing: the bureau, its customers, or both?
Discuss your requirements
If you need recordings of additive post-processing, describe the processes, tasks and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings; SourceX follows up about potential sources, and a request does not guarantee a match. Start from the humanoid training data guide or the manufacturing and assembly data hub, or read about first-person video of skilled manual work and the physical-world path for new recordings.
Describe the workplace demonstration data your robotics project needs
Sources
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — method reference for visuomotor policies trained on observation-and-action demonstrations; does not cover this setting.