Building a $300 Computer Vision Robot: A Hands-On Introduction to Fetch
A maker shares the $300 computer vision robot Fetch, built on Edge Impulse and 3D printing for STEM education.
A maker showcased Fetch, a computer vision robot on Reddit, positioned as an entry-level to intermediate STEM project costing about $300. Fetch uses Edge Impulse to train a two-class object detection model, completing the full pipeline from data collection to edge deployment. Most of its body is 3D printed with snap fits and slide rails to reduce screws. It integrates ElevenLabs voice technology for pre-recorded phrases and features movable antenna ears to express emotions. The author plans to release code, print files, schematics, and an assembly manual, currently under review at the Arduino Project Hub.
A maker shared a computer vision robot called Fetch on Reddit, positioned as an entry-level to intermediate robotics STEM project with a total cost of around $300.
Project Overview: A Low-Cost, Reproducible STEM Teaching Tool
At the core of Fetch is a two-class object detection model trained using the Edge Impulse platform. This means it can recognize two specific types of objects. Although the task is simple, it walks through the complete pipeline of "data collection—model training—edge deployment." The author emphasizes that the design deliberately kept costs down, with the final bill of materials coming in at around $300—a quite accessible threshold for enthusiasts who want to get hands-on with the combination of computer vision and robotics.
It's worth noting that Fetch is the author's first autonomous robot. Building the mechanical structure, electronic hardware, and software logic from scratch is itself a complete learning cycle.
Edge Impulse is a development platform for embedded machine learning, offering a one-stop toolchain from data collection, labeling, and model training to deployment on edge devices. It supports multiple modalities such as images, sound, and sensor data, and optimizes model size and inference speed for resource-constrained microcontrollers (such as Arduino, Raspberry Pi, and ESP32). For beginners, Edge Impulse's graphical interface and pre-built templates greatly lower the barrier to machine learning—you can complete end-to-end model development without deeply mastering deep learning frameworks. The Fetch project uses this platform to train its two-class object detection model, reflecting the application of the "low-code/no-code" AI development trend in robotics education.
Hardware Design: 3D Printing and Snap-Fit Assembly
Most of Fetch's body is manufactured through 3D printing. To reduce assembly complexity, the author heavily uses snap-fit and slide-rail mating in the structural design, minimizing the use of fasteners such as screws. This approach not only reduces the number of parts but also lowers the assembly barrier—for educational scenarios, students can finish building faster and focus their energy on programming and debugging.
The specific print files, circuit schematics, and bill of materials have not yet been released. The author says they have been submitted to the Arduino Project Hub and are awaiting review, with links to be added once approved.
Snap-fit assembly is a mechanical connection method that requires no additional fasteners such as screws or glue, achieving engagement through the elastic deformation of the parts themselves. In 3D printing, designing snap fits requires considering material toughness (e.g., PLA is brittle, while PETG is better suited for snap fits), print layer orientation, and tolerance control (typically allowing a 0.2–0.4 mm gap). Slide-rail mating achieves linear motion or positioning through a guide rail and slider, commonly found in structures that need frequent opening or adjustment, such as drawers and battery compartments. This type of design not only reduces part count and assembly time but also lowers the failure rate caused by lost screws or improper tightening force, making it especially suitable for teaching scenarios where students repeatedly disassemble and reassemble.
Interaction and Behavior: Sound and Expressive Movements
Beyond visual perception, Fetch also has some interactive expressive capability. It uses ElevenLabs' voice technology to play pre-recorded phrases. At the same time, the author designed movable parts such as "antenna ears" to express emotions or states. These details make the robot no longer just a cold actuator but give it a certain anthropomorphic quality, which is more likely to spark learners' interest for STEM demonstrations and teaching.
ElevenLabs is a company focused on AI speech synthesis, known for its text-to-speech (TTS) service that delivers highly natural and emotionally expressive output. In robotics projects, developers can pre-synthesize text into audio files and store them locally on the device or generate them in real time via the cloud, then play them through a speaker. Compared with traditional recorded playback, TTS solutions make it easy to quickly modify lines, support multiple languages, and offer a rich selection of voices. For educational robots like Fetch, using ElevenLabs to play pre-recorded phrases ensures voice quality while avoiding the computational burden of running complex speech synthesis models on embedded devices.
Open Resources Worth Watching
The author mentions that the complete code, print files, schematics, parts list, and assembly manual will all be released together with the Arduino Project Hub page. In addition, he has made a YouTube video for demonstration. For enthusiasts who want to reproduce or draw inspiration from the project, these resources will greatly lower the difficulty of getting started.
Detailed documentation for the project is still under review, but the content already shown is enough to outline a clear technical path: use Edge Impulse to train a lightweight vision model, use 3D printing and snap-fit structures to control hardware costs, and use voice and motion design to enhance interactivity. For an entry-level robotics project, this is a reference case with a complete structure and controllable cost.
One-line Summary
A maker shares the $300 computer vision robot Fetch, built on Edge Impulse and 3D printing, aimed at STEM education.
Paragraph Summary
A maker showcased a computer vision robot called Fetch on Reddit, positioned as an entry-level to intermediate STEM project with a total cost of about $300. Fetch uses the Edge Impulse platform to train a two-class object detection model, completing the full pipeline from data collection to edge deployment. On the hardware side, most of the body is 3D printed, using snap fits and slide rails to reduce screws and lower the assembly barrier. For interaction, it integrates ElevenLabs voice technology to play pre-recorded phrases and features movable parts such as antenna ears to express emotions. The author plans to release the code, print files, schematics, parts list, and assembly manual, currently under review at the Arduino Project Hub. The project provides a reference case with a complete structure and controllable cost for entry-level robotics education.
Related articles
VERS: Editing Activation Explanations to Halve Reasoning Model Tokens
VERS guides reasoning models by editing activation explanations, cutting 45.8% of tokens on GSM8K with only a 0.82 pp accuracy drop and transferring to larger models, but losing accuracy on competition problems.
Finding Errors Is Not Enough: Selective Model Escalation for Entity-Targeted Sentiment Analysis
This study explores selective model escalation for entity-targeted sentiment analysis, comparing data collection ordering, uncertainty-driven re-examination, and model routing—finding that minority-class priority only helps at very small data volumes, the oracle upper bound far exceeds actual gains, and escalation must be tailored to each dataset.
WAS-PO: Using Whitened Advantages to Solve the Correlation Trap in Multi-Reward Reinforcement Learning
WAS-PO paper proposes whitening advantages to correct correlation bias in multi-reward RL: when rewards are correlated, naive weighted summation effectively applies Σ^{1/2}w instead of w. Experiments show training reward improves steadily, but held-out accuracy does not improve in sync, and under two rewards it actually falls behind.