They're early demos, so expect glitches, queues and limits. Notes on each one's limits are below.
Odyssey-3type a sentence, then walk around inside it
experience.odyssey.systems · by Odyssey (odyssey.systems) · research preview since Oct 2026
You describe a place, and an AI world model streams it to you as live video that you steer with W A S D. It isn't a game engine and nobody filmed it: every frame is generated as you move.
- My test: "A red fox walking through a snowy pine forest at dawn". The world was playing about 15 seconds after I pressed Enter, and holding W walked the camera through the trees.
- Then I typed "night falls and the northern lights appear" into Prompt to shift the simulation, and the sky changed within seconds.
- Limits: a session lasts 5 minutes, the picture is low resolution (832×480 for the base model) and it warps and drifts the longer you go.
- Its terms say you own what you make, but the free app is for personal, non-commercial use.
- For comparison, Google's similar world model (Project Genie) sits in its top Google AI Ultra plan.
MotionBricksNVIDIA's motion AI, running in your browser tab
huggingface.co/spaces/townbox/motionbricks-playground · model: NVIDIA (SIGGRAPH 2026, github.com/NVlabs/motionbricks, 975 ★)
A 195M-parameter model generates a humanoid robot's (Unitree G1) movement live, about once a second, while you steer. It runs on your own machine through ONNX Runtime Web, so nothing is sent to a GPU server.
- Click the floor and the robot walks there; double-click to run.
- Try the styles: Zombie, Vogue, Injured Leg, Combat.
- Walk it into the wall or the bar and it vaults, ducks or sits on its own.
- The playground is a community browser port (by townbox) of NVIDIA's official model. Model: NVIDIA Open Model Agreement. Code: Apache-2.0.
FlyDronesfruit-fly neurons fly a virtual drone
spikecalls.github.io/FlyDrones · github.com/SpikeCalls/FlyDrones · MIT · 274 ★
A spiking neural network built from fruit-fly cell types steers a 3D drone in a virtual bedroom. Camera → fly eyes → brain → descending neurons → throttle.
- Press Space for autoplay, S to swat, or 1–6 for hand gestures. "Use my hand" uses your webcam.
- In the video, a hand rushes at the camera, the fly's giant-fibre escape circuit fires and the drone darts away.
- Be clear on what it is: the browser runs MiniFly, an 850-neuron stand-in with real fly cell types. The full ~166,000-neuron male fly connectome loads only in the Python package. It's a simulation, not a real flight.
TripoSplatone photo becomes a 3D object
huggingface.co/spaces/VAST-AI/TripoSplat · github.com/VAST-AI-Research/TripoSplat · MIT · 1.4K ★
Upload one picture and it builds a 3D Gaussian splat (262,144 of them) that you can orbit and download as .ply.
- My test: a public-domain (CC0) photo of a teapot by Ōtagaki Rengetsu, c. 1875, Cincinnati Art Museum. The server time was 11.1 s, about 16 s door to door.
- The back of the object isn't in the photo, so the model guesses it.
- Limit: it runs on Hugging Face's free GPUs. Visitors get a small daily quota (I ran out after a few runs). A free Hugging Face account gets more.
Fuglerammea garden mic that draws every bird it hears
fugleramme.arnegiacomo.dev (live) · github.com/arnegiacomo/fugleramme · MIT · 4K ★
A microphone in Arne Giacomo's garden in Bergen, Norway, listens all day. BirdNET (Cornell Lab and TU Chemnitz, via BirdNET-Go) names each bird it hears, and the frame shows it as a hand-cut illustration from real 1800s natural-history plates (von Wright, Gould and others). The art is real, not AI.
- The live page is his garden, right now. It changed overnight between my two visits.
- Build your own: a Raspberry Pi or any homelab box, a USB mic and BirdNET-Go. The repo has a Docker setup and runs on an e-ink panel, a TV or any screen.
- Licences: the code is MIT and the cut-outs are CC BY-SA 4.0. BirdNET-Go is CC BY-NC-SA 4.0 (non-commercial).
More free guides: pranavinnovates.com