Device
Gotchi
A small AI companion built on the SenseCAP Watcher W1-B. It listens and shows your pet; Gotchi Local on your computer handles speech through the FrontierOS backend, and can download Whisper.cpp and Piper as offline fallbacks.
Hardware
The devices, robots and compute boards physical AI runs on. Compute tiers are estimates from parameter counts, not measured results.
Devices & robots
Repos listed here document use with the device or robot.
Robot
SO-100 / SO-101 arm
A low-cost open-source 6-DoF arm designed by The Robot Studio with Hugging Face, supported natively by LeRobot for teleoperation, recording and training.
Robot
ALOHA bimanual
A low-cost two-arm teleoperation rig from Stanford. Its simulated tasks are the standard starting point for bimanual policies such as ACT.
Robot
Reachy Mini
An open-source expressive desk robot from Pollen Robotics and Hugging Face, programmable in Python and extended with Spaces apps.
Compute
Model fits are estimated from parameter counts, not benchmarks.
ComputeMicrocontrollerEstimated tier: ESP32-S3 class, a few MB of RAM
ESP32-S3
A dual-core Wi-Fi and Bluetooth microcontroller with vector instructions, suited to wake words, keyword spotting and small sensor models.
ComputeSingle-board computerEstimated tier: Raspberry Pi 5 class CPU
Raspberry Pi 5
The go-to single-board computer for robots and voice devices that need small speech, detection and depth models without a GPU.
ComputeEdge GPUEstimated tier: Jetson Orin class, 8–16 GB shared memory
NVIDIA Jetson Orin
Embedded GPU modules for robots, from Orin Nano to AGX Orin, running CUDA and TensorRT on the robot itself.
ComputeLaptopEstimated tier: Recent laptop with 16 GB or more
Laptop host
The computer a device pairs with. Gotchi Local runs here, and most speech stacks and small multimodal models fit comfortably.
Compute24 GB GPUEstimated tier: Desktop GPU with about 24 GB of memory
GPU workstation
A desktop with a 24 GB class GPU for fine-tuning policies and running larger VLAs.