Nikhil Abraham

Nikhil Abraham is a co-founder of CloudChef, where his work centers on translating chefs’ knowledge into software and robotic kitchen labor. Speaking as its CEO at AI Engineer World’s Fair 2025, he described robots that learn recipes from a single expert demonstration, combining physical motion skills with thermal and visual perception. He also appeared in a May 31, 2025 Latent Space kitchen-tour interview.[2][1]

In his 2025 conference talk, Abraham explained CloudChef’s collective engineering approach: fine-tune robot foundation models for actions such as picking and stirring, use teleoperation for edge cases, and represent recipes as state machines grounded in cooking-specific thermal and visual embeddings. His example was onions that might need seven minutes of sautéing one day and nine the next. In the demonstration he described, the robot compared current onion browning with the chef’s recorded browning, aiming to reproduce the cooking state rather than simply repeat a duration.[2]

Abraham said the team installed sensors in active commercial kitchens and collected data representing hundreds of thousands of cooked meals across recipes, cuisines, and seasons. It combined that private collection with public data and self-supervised models. He reported that CloudChef’s perception model outperformed expert chefs, Gemini 2.5, and o3 in a cooking-progress estimation evaluation spanning more than 1,000 recipes. The comparison concerned identifying a dish’s cooking state from live data, an expert demonstration, and a text recipe; it did not establish general superiority at cooking, and he supplied no numerical accuracy or complete evaluation protocol.[2]

Abraham distinguished perception, movement speed, and finished-meal taste as separate measures of success. Recorded cooking sequences supplied supervision for intermediate progress estimates. Physical tasks were compared with human motion speed, while in-house blind taste tests compared a chef’s reference dish with repeated robot recreations. He reported favorable taste-test results without giving sample sizes or scoring details.[2]

Abraham favored two arms on a wheeled base over humanoid robots, citing humanoids’ cost, maintenance, and reliability problems. He presented hourly robot labor at about $12 as the commercial proposition in 2025. Line cooking was the initial priority: he estimated it accounted for about half of kitchen labor costs and said the team wanted to focus on tasks that directly affected food quality. Dishwashing was outside the near-term scope; chopping remained on the roadmap because the necessary task data had not yet been collected.[2]

The system Abraham described remained dependent on people. He reported approximately 95% autonomous and 5% teleoperated operation, without specifying the denominator or observation period. Humans still prepared ingredient measurements in some deployments and replenished supplies when the robot alerted them. Most motions, he said, ran at 80–95% of human speed; he attributed the limitation partly to slow teleoperated training demonstrations and proposed reinforcement learning as a possible improvement.[2]

Abraham argued that faster cooking processes could sometimes compensate for slower robot movements, citing instances of a chef’s 20-minute process being shortened to 14 minutes. He qualified recipe optimization as experimental: it worked better in cuisines where thermodynamic modeling was straightforward and had not generalized across all cuisines.[2]

The robotic system followed an earlier division of labor. In a February 6, 2023 interview with The Spoon, Abraham described CloudChef’s founding idea as “record and playback taste.” Sensors and cameras captured temperatures, moisture loss, and onion browning, encoding a chef’s process into a machine-readable playback file. Software controlled appliances and guided cooking decisions, while people moved ingredients, weighed food, and plated meals.[4]

Abraham’s explanation emphasized the limits of robotic manipulation: “If you tell a human to scrape stirred rice from the bottom of a pan, it’s pretty intuitive. Most humans wouldn’t have a problem doing that. But teaching that to a robot takes time.” The culinary knowledge came from participating chefs, including Srijith Gopinathan, Thomas Zacharias, and Manjit Gill, who recorded recipes; the platform paid chefs royalties when their recipes were made. This earlier capture-and-playback work supplies context for the later demonstration-learning robots, which took on physical cooking tasks while retaining human assistance.[4][2]