Short answer
Commercial recipes and kitchen prep procedures give a robot team ingredients, quantities, timing, sequence and holding rules for a food task. They are written task specifications, not demonstrations: they say what the dish needs, not how a hand portions, folds or scrapes. Pair them with video or robot data if you need motion, and treat the recipe as the plan and scoring reference.
The buyer's problem
Public recipe text is abundant, but it is written for home cooks and rarely says how much, how long or how consistently. A restaurant group, commissary or central kitchen writes prep documentation for staff who must produce the same output across shifts and sites. These documents specify yields, portion weights, holding times, labeling and station layout in ways consumer recipes do not.
For a humanoid or mobile manipulator, those specifications can define the goal state of a task, order the steps for a planner, and give an evaluator an objective target. They do not teach the robot to cut, whisk or plate. A team planning food-handling work needs to know which part of the stack the documents feed.
What published video work tells you about recipe tasks
Recipe tasks are among the most studied procedural domains. YouCook2 contains 2,000 videos covering 89 recipes totaling 176 hours, with procedure segments annotated with temporal boundaries and descriptions (Zhou et al., 2017). CaptainCook4D recorded 384 egocentric recordings (94.5 hours) of recipes in real kitchens, with normal and intentionally erroneous executions, 5.3K step annotations and 10K fine-grained action annotations (Peddi et al., 2023).
These are research datasets of recipe-following by people, and they show two things relevant to a buyer. Step segmentation of recipe video is a defined task, and errors in recipe execution (wrong order, missing ingredient, wrong measurement) can be annotated and benchmarked. Neither dataset contains robot actions. Commercial kitchen documents can pair with video the same way, but your own capture and labeling will be needed.
What commercial recipes can support, and what they cannot
| Use | What the document contributes | What else is needed |
|---|---|---|
| Goal specification | Final product, yield, portion size, presentation | Perception that can judge portion and presentation |
| Planning and sequencing | Ordered steps, parallel prep, timing | Grounding in the robot's skills and the kitchen layout |
| Quantity reasoning | Weights, volumes, counts | Scales or sensors; handling of real ingredient variation |
| Error detection | Rules that define a wrong or unsafe result | Video of mistakes; see food-safety procedures |
| Language grounding | Terse kitchen vocabulary linked to ingredients and tools | Aligned video for visual grounding |
What they cannot give you: knife trajectories, force when pressing dough, how a sauce looks when it is ready, or how an experienced cook adjusts when the produce is different. Those live in practice, discussed in tacit knowledge gaps in SOPs.
Food safety rules sit around the recipe
Kitchen procedures in US retail food service sit inside a regulatory frame. The FDA Food Code is a model code that provides guidance to state and local authorities and retailers, and adoption is voluntary by jurisdiction (FDA, 2022). A recipe document may therefore carry hold temperatures, cooling steps or labeling rules that originate in a local code rather than the business's own preference.
For a buyer, this means two things. First, safety constraints embedded in a recipe may differ by jurisdiction, so check which rules the supplier's documents reflect. Second, such steps should be tagged as constraints rather than optional technique; see safety-critical steps in procedures. This is not legal or food-safety advice.
Illustrative package: structured prep procedures
# Illustrative only - synthetic example, not an available dataset
package: kitchen_prep_procedures_v0
source: central-kitchen documents from one operator (several sites)
documents:
recipes: structured entries (ingredients, yield, method steps)
prep_cards: station-level cards with par levels and hold times
images: plated-reference photos, one per recipe
fields_per_step:
step_id: rcp_031_s04
verb: "fold"
objects: [batter, bowl, spatula]
quantity: {value: 450, unit: g} # synthetic
duration_s: 90
condition: "until no dry flour visible"
tool: spatula
critical_to_safety: false
hold_rule: null
excluded: motion, force, video, taste or texture judgments
rights: internal evaluation and training of planning models; brand names removed; no redistribution
The condition field holds the most useful and the least precise content. "Until golden" or "until combined" are perceptual end states; a team building success detectors can use them as the reason to collect images at the transition, but cannot treat them as measurements.
Preparing the documents
Recipe sheets and prep cards come as spreadsheets, PDFs, POS-linked databases, laminated photos and training-app content. Preparation typically includes splitting ingredients from method, normalizing units, extracting durations and tolerances (parameters and tolerances extraction), and turning prose methods into ordered steps. Branching steps, such as "if dough is sticky add flour", need to be preserved as decision points; see branching procedures and decision points.
If you want to align documents with video of the same recipes, SOP-to-video step alignment describes the labeling. Turning steps into planner-ready structure is covered in task graphs from procedures.
Rights and provenance
Signature recipes are often the core of a food business. Many operators will not license them, or will license only a subset: prep cards but not formulas, or station layouts but not sauce ratios. Ingredient quantities can be commercially sensitive even when techniques are standard. Ask who created each recipe (chef, franchisor, supplier), whether the business can license it, and whether it contains trademarked names or third-party supplier specifications.
Redacting quantities defeats a use case that depends on quantity reasoning, so decide up front whether you need real quantities, scaled quantities or synthetic ones. The redaction page, redacting SOPs, covers the trade-offs.
Limitations
- Food is variable. A recipe assumes an ingredient state that real deliveries seldom match.
- Written prep rarely documents tool handling, so grasp and cutting skills need other sources.
- Documents may describe a standard that cooks routinely adapt; see procedural drift.
- Kitchen vocabulary differs by region and operator, so alias tables are often required.
Questions to ask suppliers
- Which documents are controlled and kept current, and which are informal station notes?
- Do recipes list weights and times, or only volumes and "to taste" instructions?
- Which items come from franchisors, brand owners or ingredient suppliers?
- Are there reference photos of finished items and common defects?
- Which jurisdiction's food-safety rules do the documents follow?
- Can the business license methods without the proprietary ratios?
- Are the same recipes filmed or photographed anywhere, such as training content?
- What languages are used on kitchen cards, and are translations controlled?
Discuss your requirements
If you are scoping a food-preparation task, describe the recipes, formats 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. A request does not guarantee a match; SourceX follows up about potential sources.
See also the humanoid training data guide and the SOPs and procedural data hub. The catalog entry for robotics and embodied AI data shows how procedural documents fit into wider requests.
Discuss your humanoid training-data requirements with SourceX
Sources
- Towards Automatic Learning of Procedures from Web Instructional Videos — YouCook2 pairs recipe videos with temporally bounded procedure segments and descriptions.
- CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities — CaptainCook4D records recipes in real kitchens including intentional errors, with step and action annotations.
- Food Code 2022 — FDA Food Code is a voluntary national model for retail food safety adopted by state and local authorities.