About This Project

Could an affordable robot reliably pick up and deliver bites of food? We hypothesize that an SO-101 arm using camera-based bite selection and pickup verification can successfully deliver at least 80% of bites in controlled tests. We will run 150 trials across five food types and measure acquisition, transfer, spillage, timing, and failure detection. Tests will use a soft mannequin, not people. We will openly share the design, code, data, failures, and results.

Ask the Scientists

What is the context of this research?

Eating independently is easy to take for granted. People with limited arm or hand movement may need help throughout a meal, reducing privacy and placing demands on caregivers. Prior research shows that robots can learn different skewering strategies for foods with different physical properties, while combining vision and touch can improve acquisition of unfamiliar foods. Other work reports that a general-purpose utensil can acquire varied foods at 80% or greater success. We will test whether these capabilities can be reproduced on the inexpensive, open-source LeRobot SO-101. We hypothesize that it can achieve at least 80% end-to-end success in four of five controlled food conditions when camera-based bite selection is paired with pickup verification.

What is the significance of this project?

A practical, affordable feeding assistant could help people with limited upper-body mobility participate in meals with greater independence and dignity. Before such a system can be tested with people, however, it must reliably handle everyday variation: foods move, deform, slip, and may not remain on a utensil. This project focuses on those basic technical challenges and on detecting failures rather than assuming every attempt succeeds. The results will show what a low-cost robot can and cannot do, identify the main barriers to safe operation, and provide open designs, code, data, and evaluation methods that other researchers can reproduce and improve.

What are the goals of the project?

I will evaluate the Lerobot SO-101, which is low cost, compact open-source tabletop robot arm. There will be 150 trials: 30 each with banana, strawberry, tofu, melon, and pasta, representing soft, slippery, deformable, firm, and loose foods. Trial order will be randomized. A trial succeeds when the robot acquires a bite and delivers it to the fixed mannequin target without dropping it. We will also record acquisition success, transfer success, spillage, cycle time, and whether the system correctly detects a failed pickup. Our Phase 1 benchmark is at least 80% end-to-end success in four of five food conditions, at least 90% failure-detection accuracy, and no more than 10% drops or spills. These are research thresholds for controlled mannequin testing, not evidence of safety for human use.