Commercial Kitchen Food Prep Demonstrations: Cutting, Portioning and Plating

The deployment problem

A commercial kitchen is not a home kitchen at larger scale. Recipes are standardized and written as specifications, prep happens in batches, each station has its own tools, and food-safety rules govern which board, glove and container touches which ingredient. Work runs at service pace, with inspectors checking temperatures and labeling.

The best-known public kitchen datasets were recorded in homes or taken from the web. A team deploying a humanoid in prep or plating needs to know which public data transfers, what commercial operators can add, and which layers neither can supply.

What public kitchen datasets cover

Three widely used datasets set the baseline:

• EPIC-KITCHENS-100 recorded 100 hours of head-mounted video across 45 kitchens, annotated with 90K actions (Damen et al., 2020). The kitchens are domestic and the activity unscripted.
• YouCook2 collected 2,000 web videos covering 89 recipes, 176 hours in total, with procedure segments marked by temporal boundaries and descriptions (Zhou et al., 2017). The video is edited for viewers, not recorded for robots.
• CaptainCook4D recorded 384 egocentric sessions, 94.5 hours, of people following recipes in real kitchens, including intentionally erroneous executions, with 5.3K step annotations; it is released under Apache 2.0 (Peddi et al., 2023).

These are good material for cooking step recognition, action recognition and error detection. They do not show commercial specifications, batch volumes, multi-station handoffs, color-coded food-safety practice or service pace. That gap is what commercial suppliers can fill.

Task anatomy

Around these sit mise en place, glove changes, handwashing and sanitizing surfaces between tasks. A humanoid that cuts well but skips a board change has failed the task.

Which data helps each part

Cutting. Human video shows grip, stroke pattern, how the off hand guides the food and the sequence from whole item to diced product. It does not show the force through the blade or how a tomato skin resists before it splits; that needs robot, teleoperation or instrumented-tool data, as set out on when video is not enough. Recognizing ingredients and their state (whole, peeled, diced) is a perception task covered in food item recognition.

Portioning. This is where business systems help most. Many kitchens weigh portions or check them against a tool size, and some log the result. Linking each portion event in video to its measured weight turns footage into labeled examples of in-spec and out-of-spec portions; aligning video with system logs covers the join.

Plating. Build cards with reference photos are ready-made success criteria, and plating video shows component order and hand placement. Line-style assembly at volume is treated on food assembly line hand tasks.

Procedures. Recipes, prep lists and build cards supply the step vocabulary and parameters; see recipes and food prep procedures. Utensil and equipment identity is a separate recognition need, covered in kitchen equipment and utensils.

Errors. Normal executions are useful even without labeled mistakes: EgoPER trains its procedural error detector only on error-free videos (Lee et al., 2024). Real errors from a commercial line are still needed to test such a detector.

Capture and labels

Two views cover most prep work: a head-mounted camera for hands, knife and food, and a top-down camera over the board or plate that shows cut sizes and layout. Plan for steam, glare from steel surfaces and lens fogging near ovens. Mount equipment so it meets the site's hygiene rules, and agree where cameras may sit near open food.

# Illustrative only — synthetic label record for one portioning episode
episode_id: prep_site031_0552
station: cold_line_2
spec: {card_id: BC-0187, item: grain_bowl, components: [base, protein, veg_mix, sauce, garnish]}
operator: cook_C04 (experience: 3 yr)
cameras: [head_rgb_1080p60, overhead_rgb_1080p30]
events:
  - {label: portion_base,    tool: scoop_8oz,   target_g: 180, measured_g: 186, in_spec: true}
  - {label: portion_protein, tool: scale,       target_g: 110, measured_g: 97,  in_spec: false}
  - {label: sauce,           tool: squeeze_btl, pattern: zigzag}
  - {label: garnish,         placement: "2 o'clock", matches_photo: true}
food_safety: {glove_change: true, board_color: green, allergen_flags: [sesame]}
outcome: {matches_build_card: false, corrected: true}
excluded: robot actions, knife or scoop forces, food texture measurements

Record the spec version with every episode; build cards change with menus, and an out-of-date card turns correct plates into apparent errors.

Supplier types

• Multi-unit restaurant groups and quick-service chains. Standard build cards, training video and station layouts repeated across many sites.
• Institutional kitchens. Hospital, university and corporate dining operators doing batch prep at volume.
• Commissaries and central kitchens. High-volume cutting and portioning for many outlets.
• Prepared-meal producers. Portioning and assembly lines with weight checks.
• Culinary schools. Recorded practicals with instructor assessment, subject to learner consent.

Recipes and build cards are often treated as confidential, and chain training video may belong to a franchisor or a production vendor rather than the operator. Ask who can license each asset and what must be redacted. Workers on camera need releases that cover AI training; customers rarely appear in back-of-house footage, but dining-room shots should be excluded. New recording follows the site's separate physical-world path for hands-on work, and the catalog entry for first-person video of skilled manual work describes the related asset type.

Limits for this task

• No feel. Blade force, food resistance and slip are invisible in video.
• Ingredient variation. Size, ripeness and temperature change how food cuts and portions; record them where possible.
• Gloves and occlusion. Gloved hands, knives and the food itself hide grip details.
• Speed and blur. Experienced cooks cut faster than a robot will; raise frame rate for knife work.
• Domestic data mismatch. Home-kitchen datasets lack commercial specs, volume and station handoffs.
• Safety. Knife handling by a humanoid raises safety questions that data alone does not answer.

Questions to ask suppliers

1. Which prep tasks are done by hand, and which by slicers, processors or portioning machines?
2. Are recipes and build cards current, versioned and illustrated with reference photos?
3. Are portion weights or temperatures logged, and can logs be linked to timestamps?
4. Which stations and shifts could be recorded without disrupting service or food safety?
5. Does training video belong to the operator, a franchisor or a production vendor?
6. Which recipes or specs are confidential, and what would need redacting?
7. Are mistakes and corrections ever visible, or is footage edited to show correct practice?
8. Do cooks' releases cover AI training and third-party licensing?

Discuss your requirements

If you are planning a kitchen deployment, describe the prep tasks, specs and capture you need 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. It reviews each requirement and follows up about potential sources; a request does not guarantee a match. See the workplace task demonstration hub for related tasks and the humanoid training data guide for the full section.

Describe the workplace demonstration data your robotics project needs

Sources

• Rescaling Egocentric Vision — head-mounted kitchen video across many domestic kitchens.
• Towards Automatic Learning of Procedures from Web Instructional Videos — recipe videos with annotated procedure segments.
• CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities — recipe recordings with normal and erroneous executions.
• Error Detection in Egocentric Procedural Task Videos — error detection trained on error-free procedural video.