Community
Discussions for people building robots and AI devices: what to train, what to run it on, and what you learned along the way.
All posts
Welcome: show us what you're building, and tell us what to index next
FrontierOS Hub collects open models, datasets and Spaces for robots and AI devices in one place. Every repo links to its home on Hugging Face; nothing is hosted or mirrored here yet.
Fine-tuning SmolVLA on an SO-101: what's your recipe?
The SO-101 pick-and-place set is a good template: about fifty episodes with a side and a top-down camera. Most people seem to start from the SmolVLA base checkpoint.
Which offline voices should Gotchi Local ship next?
Today Gotchi Local can download Whisper.cpp for transcription and a Piper voice for speech, both from Hugging Face, as offline fallbacks for when it is not using the FrontierOS backend.
What's the lowest-latency fully local voice loop you've built?
A typical cascade is voice activity detection, a turn detector, streaming speech-to-text, a small language model and a fast text-to-speech voice.
pi0, pi0-FAST or pi0.5: which base do you fine-tune first?
LeRobot now ships ports of three Physical Intelligence checkpoints. pi0 uses flow matching, pi0-FAST predicts compressed action tokens, and pi0.5 adds co-training aimed at unfamiliar scenes.
What actually runs well on a Raspberry Pi 5?
Our compute tiers are estimates based on parameter counts. Real numbers beat estimates.
Open humanoid data is still scarce. What are you training on?
The permissive options are few: Humanoid Everyday for real teleoperation, NVIDIA's GR00T simulation sets, and large egocentric human video as a proxy.
Check your teleop data before you train
Step through a few episodes in the LeRobot dataset visualizer before spending GPU hours.
How Gotchi uses models today
The Watcher listens and shows your pet. Heavy models do not run on the device itself.
Full-duplex or cascaded voice for robots?
Full-duplex models such as Moshi can listen while they talk, which makes interruptions and backchannels feel natural. Cascaded stacks are easier to debug and swap piece by piece.
Open-vocabulary perception on a Jetson Orin: what's your pipeline?
Grounding DINO for boxes followed by SAM 2.1 for masks is a common language-driven pick-and-place front end.
Which cross-embodiment mix would you pretrain on?
Open X-Embodiment, DROID and BridgeData are the usual building blocks, with LIBERO as the common simulated benchmark.
Are video world models ready for robot planning?
V-JEPA 2 plans in a learned latent space, while Cosmos generates future video that can stand in for rollouts or synthetic training data.