Liquid Pouring and Dispensing Demonstrations

Task anatomy

A pour looks simple and splits into phases that fail in different ways.

1. Approach and grasp. Hold the source container so the opening is clear; handle the lid, spout or handle.
2. Align. Position the spout over the target; judge distance and height.
3. Tilt. Rotate until flow starts; the angle at which it starts depends on liquid level, viscosity and container shape.
4. Control flow. Adjust tilt as the level drops; watch the target's level rise.
5. Stop. Reduce tilt before the target is full, accounting for the stream already in flight; avoid drips.
6. Return and place. Right the container, wipe or set down.

Dispensing adds mechanisms: pumps, taps, pressing a dispenser lever, filling to a line, measuring with a scoop. Failure modes include overfill, spills, splashing, missing the target, dripping on the rim, and a drained source. Tasks vary by container (bottle, pitcher, carafe, flask, beaker), liquid (water, milk, syrup, oil, reagent), and target (cup, bowl, pot, vial).

What video can support

Related datasets show what recording kitchens and labs produces. EPIC-KITCHENS-100 collected 100 hours of head-mounted video in 45 kitchens, with narration first and then transcription, parsing and temporal segmentation into 90K actions (Damen et al., 2020). CaptainCook4D recorded 384 egocentric recordings (94.5 hours) of recipes in real kitchens, including both normal and intentionally erroneous executions, under IRB approval with written consent (Peddi et al., 2023). FineBio recorded 32 participants performing mock biological experiments (14.5 hours), annotating protocols, steps, atomic operations, object locations and manipulation states (Yagi et al., 2024).

These sets support recognition of pouring as an action, step localization and error recognition. None is a source of measured liquid volumes or flow rates, so they speak to the "what happened" half of the problem. For the "how much" half, you need instrumentation. See laboratory bench tasks and beverage preparation for neighboring task pages.

What video cannot support

• Volume and flow rate. Clear liquids in clear containers are hard to see. Reflections and meniscus effects confuse level estimation.
• Liquid properties. Viscosity, temperature and surface tension are not in pixels.
• Container weight and balance. A full pitcher feels different from an empty one; video shows the arm, not the load.
• Robot control. Tilt rates and trajectories from a human wrist do not translate to a humanoid's wrist limits. See the human-to-humanoid embodiment gap.
• Safety constraints. Hot or hazardous liquids create risks that are rarely captured in routine footage.

Simulators that model liquids, such as the one behind BEHAVIOR-1K, can generate pouring scenes, but they are a separate data type from recordings of people at work; see when video is not enough for how to decide what else to add.

Variation that matters

Pouring is sensitive to variables a casual recording will hold constant by accident. Plan the variation table before the first shoot.

Fill level changes the tilt needed for flow onset, so record the source container at full, half and nearly empty. Container shape changes how the stream leaves: a pitcher with a spout behaves differently from a wide-mouth jar or a narrow-neck bottle. Target size changes the margin for error, and a small vial is much less forgiving than a mixing bowl. Liquid viscosity changes both flow and drips; water and syrup need different strategies. Operators differ in handedness and technique, so include several.

Most important for evaluation, hold out some combinations entirely. If every pour in training uses the same pitcher, a held-out pitcher is the real test. Write which containers, liquids or sites are held out into the delivery terms and keep them out of training; the principle is explained in demonstration dataset splits.

Hygiene and safety shape what you can ask for. Food service areas have rules about what may enter the kitchen, and a laboratory may not allow external cameras or a scale near work. Ask these questions first, because they limit camera placement and instrumentation more than technical factors do.

Capture specifics

Check whether a scale can sit under the target container; where it can, it is often a useful addition, because it gives a time series of mass added, from which flow onset, rate and stop can be read directly, and it provides an objective outcome label. Mark which containers and liquids were used in every clip.

Labels

Phase boundaries; tilt events (onset, adjustment, stop); fill level before and after as a fraction or measured value; spill flag with location; container and liquid type; dispensing mechanism; outcome (within tolerance, over, under, spilled). Add an "unobservable" tag for frames where the liquid level cannot be seen, rather than forcing a guess.

Illustrative specification

# Illustrative only - synthetic values
task: "pour 250 mL water from pitcher to beaker"
containers: [pitcher_1L_clear, beaker_500mL_graduated]
liquid: water_dyed_blue
capture:
  side_camera: 1080p, 60 fps, level height
  head_mount: 1080p, 30 fps
  scale: 10 Hz mass log, timecode-synced
runs: 40 per operator, volumes varied (100, 250, 400 mL)
labels:
  phases: [approach, align, tilt, flow, stop, return]
  outcome: delivered_mL (scale), tolerance_pass (+/- 15 mL, buyer-set)
  spill: boolean + area
safety: no hot or hazardous liquids

Suppliers and rights

Kitchens, beverage operations and laboratories pour daily. Restaurants and cafes bring customers and staff into frame; laboratories bring confidentiality about reagents, protocols and equipment. Ask whether the recording can be done in a training area or outside service hours, and review the commercial kitchen food prep page for permission points specific to kitchens. Check consent for every visible person; see worker consent and identifiable people.

Questions to ask suppliers

1. Can the recording include a scale or graduated container, synchronized to video?
2. Which liquids and containers can be used, and which are excluded for safety?
3. What camera angles keep both the spout and the target level visible?
4. Are overfills, spills and corrections kept and labeled?
5. How is the final volume recorded for every run?
6. Can volumes and containers be varied systematically?
7. Do protocols or recipes appear in the frame, and who owns them?
8. Who in the facility approves recording and the use of the data?

Discuss your requirements

If you are scoping pouring or dispensing data, describe the liquids, containers and measurement needs to SourceX. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match, and instrumented capture is scoped case by case. The demonstration hub and the humanoid training data guide place this task in context.

Describe the workplace demonstration data your robotics project needs

Sources

• Rescaling Egocentric Vision — EPIC-KITCHENS-100 narration-first annotation pipeline and kitchen scale.
• CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities — CaptainCook4D errors in real kitchens, consent and scale.
• FineBio: A Fine-Grained Video Dataset of Biological Experiments with Hierarchical Annotation — FineBio hierarchical annotation and step-order structure.
• BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation — BEHAVIOR-1K activities, scenes and simulator coverage of liquids and deformables.