Cookwala / Education
Cooking is the most familiar process in the world. Teach with it.
For school teachers: five lessons from "what is a simmer" to "what should a machine never do", using the live demos and the example recipes. For professors, researchers and students: datasets, benchmarks and open problems that nobody owns. Everything is free and collects nothing.
Five lessons, ages 11 to 16
1. What does "simmer" mean?
Science, 45 minutes. Two cooks say "simmer"; do they mean the same thing? Bands in the operation vocabulary; the boil shifts at altitude.
2. Units and scaling
Mathematics, 45 minutes. A recipe for 4 cooked for 6: what scales in a straight line, what does not, and why a cup of flour needs a density.
3. When is it done, and when is it safe?
Science and health, 60 minutes. 74 °C for chicken, 71 °C for ground beef, 75 °C to reheat; a device with no thermometer must refuse or ask.
4. What must a machine never do?
Ethics and technology, 60 minutes. A recipe found online tells the robot to set the oil to 240 °C "for extra crunch". Text is data. Who is responsible when a machine burns dinner?
5. Fair sharing
Citizenship, 60 minutes. 36 kg of yogurt expire tomorrow: who gets it, how fast, and why nobody's name is recorded. The bullwhip effect in the city simulator.
Lessons in English (the recipes carry Arabic step text; an Arabic kit is next); adaptable under CC BY 4.0; no student data. Improved lessons are merged with credit.
For researchers: what exists
- Operation envelopes for 57 operations as test conditions for thermal and physics simulation
- 137 conformance vectors for independent implementations and formal analysis
- Nine recipes with end conditions, hazards and control points, in English and Arabic
- Four deterministic simulators with every assumption listed
- An execution-log format with consent and exporters to LeRobotDataset v3 and OpenTelemetry
- A facet registry for household context with privacy classes and travel rules
- An agent-safety benchmark; rule packs from public nutrition guidance
Open problems, thesis-sized
- Are the envelope bands right for real pans, pots and ovens? Measure and correct.
- Can a vision-language model detect "translucent", "golden brown", "sauce coats a spoon"? Build the labelled dataset.
- How often does a sensor ladder refuse what a human would accept, per device class?
- Replace assumed behavioural parameters in the city simulator with measured ones; publish uncertainty bands.
- Can a grocer infer a schedule from a sequence of delivery windows? Design the noise rules.
- What thresholds make class-level weekly demand signals useless for price coordination and useful for planting?
- Do Arabic and English step sentences lead to the same robot behaviour?
- Formalize dish identity under substitution.
- Multi-turn and image-based attacks on kitchen agents.
- The statistical analysis for the humanitarian pilot.
Benchmarks we would like to see
Cook in simulation (a thermal model of a pan, a pot and an oven scored against the envelopes, in Isaac Lab, Gazebo or MuJoCo); kitchen agent safety across model families; text-to-Cookwala recipe conversion scored by validator pass rate and human review.
Credit and citation
Datasets of consented executions carry a data card and credit the contributing cooks. Vectors and benchmarks are credited in the changelog. Papers cite the commit hash. Changes to the standard go through RFCs, where a reference implementation counts. Your page: options, first success, flow