AIfred: Augmented Learning through Functional Robotic Embodiment at the Desk

Synthesis: Generative AI reaches students through a screen, so guidance lives in a different place from the handwritten work it is meant to support. AIfred closes that gap physically: a desk-based robotic arm with a projector at its end-effector reads the workspace with an overhead camera, generates task-relevant help with a Multimodal AI model, and projects it beside the user's paper. In a user study with 36 participants comparing AIfred against ChatGPT on a laptop, the two performed comparably while help was available on a quadratic-equation assignment (6.7 vs. 7.3/10, p = .41), but once assistance was withdrawn the AIfred group held at 7.0/10 while the ChatGPT group fell to 4.4/10 — a 60% higher short-term learning transfer score (p = .003). Three independent art and design professors ranked AIfred drawings first in 33 of 36 cases. The mechanism is visible in behavior: participants switched between paper and screen 63 times with ChatGPT and once with AIfred. The authors conclude that robot-mediated projection is a targeted aid for tasks whose guidance shares a spatial frame with the work, not a general replacement for screen-based assistance.

Key Findings

1. AIfred is a desk-based robotic arm with a mini projector at the end-effector, using an overhead camera and motion capture to place AI-generated guidance beside handwritten work.
2. In the math assignment, AIfred and ChatGPT performed comparably while assistance was available (6.7/10 vs. 7.3/10, p = .41), so co-location did not raise supported performance.
3. Once assistance was withdrawn, ChatGPT-assisted participants fell to 4.4/10 while AIfred-assisted participants held at 7.0/10, a 60% higher short-term transfer score (p = .003).
4. Three independent art and design professors ranked AIfred drawings first in 33 of 36 cases (92%), second in the remaining three, and never last (Kendall's W = .86).
5. Participants averaged 1 physical-digital context switch with AIfred versus 63 with ChatGPT, a 98% reduction, yet both groups rated the disruption as low (1.3/5 vs. 1.7/5).
6. AIfred scored higher on perceived learning support (4.5/5 vs. 3.4/5, p < .001), innovation (4.5/5 vs. 3.5/5, p < .001) and satisfaction (4.2/5 vs. 3.5/5, p = .04), with no loss in perceived productivity (p = .61).
7. Mean task completion time was higher with AIfred (349.5 s vs. 259.9 s; p = .02), which the authors read as sustained physical engagement rather than inefficiency.

A Robotic Arm That Puts Guidance on the Desk

AIfred is a robotic arm at the edge of a desk with a mini projector mounted at its end-effector. An OptiTrack motion-capture system tracks the robot base and a physical object on the desk, and an inverse-kinematics solver holds the projector at a fixed height while the tracked object lets the user move the projected content. A spatial assistance pipeline runs in three stages: workspace perception, in which an overhead camera captures the task context and a MediaPipe hand-gesture detector identifies the region being pointed at; context-aware content generation, in which a Multimodal AI language model (gemini-3.8-flash for reasoning, gemini-2.5-flash-image for image generation) produces guidance; and robot-mediated projection, which places that content alongside the work. Three modes cover math homework, image generation and drawing. In math mode the goal is Scaffolding rather than answer replacement — hints, formulas, intermediate reasoning cues or analogous examples — and the paper situates the work in spatial augmented reality and prior robotic projection platforms that did not interpret the task or embed support in ongoing desk work.

Comparing AIfred with ChatGPT at the Desk

The study used a mixed experimental design with 36 participants (17 men, 19 women, ages 19 to 50) recruited from one university campus, spanning humanistic, social-science, business and engineering backgrounds. Participants were assigned to AIfred or to ChatGPT (GPT-5.6 Luna) on a laptop, and each completed a quadratic-equation assignment drawn from five comparable equations, an image-generation task, and three drawing tasks — the second with their assigned system and the third with the opposite system in a crossover. A final math problem was solved with no assistance to measure short-term learning transfer, on average 35 minutes later. Measures pair behavior and performance with Self-Report Measures, making this a mixed-methods quantitative comparison of user experience and outcomes. Solutions were graded by four independent Gemini Flash 3.6 agents on a fixed eight-criterion rubric, anonymized and randomized so condition could not be inferred (mean inter-agent SD 0.53); drawings were ranked by three independent art and design professors (Kendall's W = .86).

What the Results Show

While assistance was available the two systems were statistically indistinguishable on the math assignment (6.7/10 vs. 7.3/10, p = .41). The difference appeared after withdrawal: ChatGPT-assisted participants dropped to 4.4/10 (a 40% decrease) while AIfred-assisted participants held at 7.0/10, a 60% higher short-term transfer score (p = .003). AIfred drawings were ranked first in 33 of 36 cases, against 3 for ChatGPT and none for the unassisted baseline. Participants averaged 1 physical-digital context switch with AIfred and 63 with ChatGPT, a 98% reduction, and reported higher perceived learning support (4.5/5 vs. 3.4/5, p < .001), innovation (4.5/5 vs. 3.5/5, p < .001) and satisfaction (4.2/5 vs. 3.5/5, p = .04), with cognitive demand also higher (3.4/5 vs. 2.9/5, p = .03); perceived productivity did not differ (p = .61), and completion was slower with AIfred (349.5 s vs. 259.9 s, p = .02).

Why Task Type Matters

The authors argue the benefit is task-dependent: symbolic content such as a quadratic equation can be read from a screen and reproduced on paper, while spatial guidance for drawing depends on direct visual alignment with the user's own lines, angles and proportions, which every context switch breaks. Co-location therefore matters most when instructions and task share a unified spatial frame. Because ChatGPT users switched 63 times yet rated the disruption as low (1.7/5 versus 1.3/5), attentional fragmentation appears to operate below subjective awareness. The paper positions robot-mediated projection as a targeted complement to screen-based assistance, not a general replacement, treats pointing as a low-friction act of Help-Seeking, and reads longer time-on-task as deeper embodied engagement with reduced passive transcription.

What this means for practice

• Instructors. Judge AI assistance by what Learners retain after it is removed, not by supported performance: the systems tied while help was available (6.7 vs. 7.3/10, p = .41), and only the AIfred group held its score once help was withdrawn (7.0 vs. 4.4/10).
• Instructors. Match the delivery channel to the task. Spatially co-located guidance paid off for drawing, where reference material must align with the learner's own marks, but added little on a symbolic quadratic-equation task.
• Educational technology developers. Treat projection as a delivery mode, not a novelty: pointing at physical work as the request gesture, and projecting guidance beside it, cut observed paper-to-screen switches from 63 to 1.
• Researchers. Measure context switching directly rather than by self-report — participants averaged 63 switches with ChatGPT and still rated the disruption low (1.7/5), so subjective ratings miss a cost that tracks lower transfer and drawing quality.

Limitations

• One site, 36 participants. The sample came from a single university campus (17 men, 19 women, ages 19 to 50), so results may not generalize beyond this setting.
• Transfer measured over minutes. Short-term learning transfer came from one follow-up problem solved on average 35 minutes after the first, with no delayed retention test.
• Lab-bound tracking. AIfred relies on an OptiTrack motion-capture system, which the authors state limits deployment outside controlled laboratory settings.
• Grading and design scope. Math solutions were graded by four AI agents and drawings ranked by three professors, and the prototype prioritizes functional projection over expressive robot movement.

Citation

Orlando, G., Groshev, M., & Castelló Ferrer, E. (2026). AIfred: Augmented Learning through Functional Robotic Embodiment at the Desk. arXiv preprint.