Short answer
Lab bench demonstrations capture small-object work such as pipetting, tube handling, plate loading and instrument operation, ideally linked to a written protocol. They can support protocol-step recognition, object-state tracking and evaluation of procedure following. They do not provide robot actions, liquid-handling forces or proof of scientific validity, and bench work is often hard to see in video.
Why bench work is its own task family
Bench work is small, precise and procedural. A protocol fixes the order of steps, while the hands perform atomic operations such as opening a tube, aspirating, dispensing and discarding a tip, whose order can vary. Objects are small, transparent, labeled by tiny text and often partly hidden by the hand or by the instrument hood.
A humanoid or mobile manipulator in a lab would need to follow a protocol, handle consumables and operate instruments. Buyers asking about "lab automation training data" are often looking for the protocol layer and the manipulation layer separately. The protocol layer is covered by laboratory SOPs; this page covers recorded human bench work.
Task anatomy
| Level | Example | What to label |
|---|---|---|
| Protocol | Prepare a dilution series | Protocol ID and version |
| Step | Add 100 microliters of diluent to wells A1-A8 | Step boundaries, step index |
| Atomic operation | Attach tip, aspirate, dispense, eject tip | Verb, manipulated object, affected object |
| Object state | Tube capped / open; plate lid on / off | State before and after |
| Location | Rack position, deck slot, hood zone | Object locations over time |
This hierarchy matches FineBio, which recorded 14.5 hours of multi-view video of 32 participants performing mock biological experiments, with hierarchical annotation of protocols, steps, atomic operations, object locations and manipulation states. In FineBio the step order is strict but the order of atomic operations is flexible, which affects how a step segmenter should be evaluated (Yagi et al., 2024).
Failure modes: tip not seated, air bubble, wrong well, a dropped cap, a tube mislabeled, cross-contamination from touching a non-sterile surface, and a skipped incubation.
What the data can support
- Protocol-step recognition. Multi-view video with step labels trains models that track which step a person is in.
- Atomic-operation and object detection. Small transparent objects and manipulated-versus-affected object labels support detection and state classification. See hand-object interaction annotation.
- Procedure-following evaluation. A held-out set of runs, some with deliberate or natural mistakes, tests whether a system detects deviation.
- Language grounding. Written steps matched to recorded operations support instruction-following.
- Layout and workflow knowledge. Where consumables sit, how racks are filled and how a hood is used.
What it cannot support
- Robot actions and fine force. Seating a tip, piercing a septum or releasing a droplet depends on sub-millimeter motion and touch. Neither is in video; see when video is not enough.
- Embodiment match. Human fingers do things a humanoid hand will not. See the embodiment gap.
- Volume and meniscus. A camera at bench scale rarely resolves liquid level. If volume matters, you need instrument logs, gravimetric records or a macro view.
- Scientific validity. Video shows the operation, not whether the assay worked.
- Real research content. Real experiments contain confidential samples, compound names and proprietary protocols; mock experiments protect those but may be less varied.
GLP and protocol records
In regulated nonclinical studies, 21 CFR Part 58 sits behind much of the written record. Its Subpart E includes Section 58.81 on standard operating procedures. Each study must have an approved written protocol stating objectives and methods, and deviations from SOPs must be authorized by the study director and recorded in the raw data (FDA, eCFR Part 58). For a buyer, that means a GLP-style lab may hold a protocol version, an SOP and a deviation record that can be linked to a video, which is a stronger provenance chain than most workplaces have. Whether that material can be licensed is a separate question for the lab and its sponsors. This is not legal advice, and a recording that is part of a study's raw data may carry retention and confidentiality obligations of its own.
Capture choices
| Choice | Why | Note |
|---|---|---|
| Multiple synchronized views | Occlusion by hands and hood | Clock sync matters; see multi-view synchronized recording |
| Macro or close-up camera | Tip, well and tube detail | Fixed focal plane |
| Head-mounted view | Operator's perspective | Hood glass and glare |
| Instrument and LIMS logs | Ground-truth for volumes and times | Needs export from instrument software |
| Gloved-hand setting | Realistic | Trackers struggle; label glove type |
Small-object handling is also a fragility problem for hardware; for general handling guidance see fragile item handling.
Illustrative bench recording spec
# Illustrative only; synthetic values
dataset: bench_tasks_v0
protocols: [serial_dilution_p01, plate_transfer_p02, centrifuge_load_p03]
sites: lab_004, lab_011
participants: operator_B01..B12; experience_level tagged
cameras: [top_down_macro, front, operator_head] # synchronized to a shared clock
annotations:
protocol_id, protocol_version
steps: index, start_s, end_s
atomic_ops: verb, manipulated_object, affected_object, start_s, end_s
states: tube_cap, plate_lid, tip_attached
errors: type, step_index, corrected: true/false
instrument_logs: dispense volumes and timestamps where available
excluded: robot actions, force, sample identity
rights: protocol text licensed or replaced with a neutral equivalent
Acceptance criteria
- Step order matches the protocol version stated for each recording.
- Atomic operations carry both manipulated and affected objects, with sampled agreement checks.
- At least one camera keeps the tip, well and hand visible during dispense in a random sample.
- Errors and corrections are preserved and flagged.
- Evaluation runs are held out by lab, operator and protocol.
Suppliers and permissions
Likely suppliers are contract research and testing labs, which run documented protocols repeatedly. Questions that matter early are about whose protocol it is (a client's or the lab's), whether samples, labels or instrument screens reveal client information, and whether staff agreed to be recorded. Customer confidentiality clauses can block licensing even when staff consent exists. Redacting screens or labels does not by itself settle those obligations; see who owns a procedure.
Questions to ask suppliers
- Which protocol versions do the recordings follow, and can the text be licensed too?
- Is the work mock, training or real client work?
- Which camera views exist, and are they synchronized?
- Are instrument or LIMS logs available for the same runs?
- Do recordings include mistakes and corrections?
- What client or sample information is visible on labels and screens?
- How many operators are recorded, and at what experience levels?
- Who at the lab can authorize licensing, and do client contracts restrict it?
Discuss your requirements
If you need bench-task data, describe the protocols, views, modalities 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. If nothing existing fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match; SourceX follows up about potential sources.
Start with the humanoid training data guide or the workplace task demonstration hub. Related: step segmentation annotation, the first-person manual-work video catalog entry and the egocentric video buying category.
Describe the workplace demonstration data your robotics project needs
Sources
- FineBio: A Fine-Grained Video Dataset of Biological Experiments with Hierarchical Annotation — Multi-view lab-protocol video with hierarchical protocol, step and atomic-operation annotation
- 21 CFR Part 58 - Good Laboratory Practice for Nonclinical Laboratory Studies — GLP rules on written protocols and recorded SOP deviations