Zero setup
Uplift Cloud
One account, nothing to install. Run the VLM on frontier models and train policies from your datasets on our GPUs.
- VLM inference for every skill
- Policy training from your episodes
- Models delivered straight back to the robot
Raven
Fully trainable desktop robot you can build at home.
Hardware
Human proportions.
Seven degrees of freedom.
Raven is built like the upper body of a person, so teaching it feels natural. Start from prebuilt motion templates, chain them into action sequences for complex tasks, or use them to create training datasets.
- 7-DOFper arm, shoulder to gripper
- Vision enableda camera in the head, looking where it works
- Desk mountarrives assembled, ready to run
- CAD-accurate URDFfor simulation and advanced training in NVIDIA Isaac Sim
Vision AI
Sees. Thinks. Acts.
The camera in Raven's head reads the scene. A vision-language model reasons about what it sees, in response to a prompt you wrote in plain English. Then the robot moves.
The Raven app
One app. Novice to model training.
Record your first motion in minutes. Turn every replay into a LeRobot-compatible training episode. Train a model on Hugging Face or Uplift Cloud and run it on the robot. All in one place, on the devices you already own.
- Android
- Windows
- macOS
- iPad**
** iPad requires an additional ESP32 module.
Teach by moving.
Imitate mode mirrors your pose onto a digital twin of Raven, joint by joint, so you see exactly what the robot will do before it moves. Switch to the real robot when you are ready.
Start at your level
Beginner, Builder, or Expert. The app grows with you, from one-tap skills to AI Gym, training datasets, and policy models.
Collect real data
Every run becomes an episode in a LeRobot-compatible dataset: both camera views and every joint, synchronized, ready to scrub through and ready to train on.
Build and train
Chain motions, prompts, and controls into a skill in Skill Designer. Train a policy from your episodes on Hugging Face with your own account, or on Uplift Cloud, and run it on the robot.
Training and Inference
Your data. Your models. Your choice of compute.
Raven records LeRobot-compatible datasets and runs vision-language models for every skill. Where that compute happens is up to you.
Inference
Local compute
Run the VLM on your own capable machine. Fully offline, your data stays on your network.
- Compact vision-language models on-device
- No internet required at runtime
- Works with the same skills and prompts
Training
Hugging Face
Push your LeRobot-compatible datasets to Hugging Face and train with your own account and your own budget.
- Standard LeRobot dataset format
- Train with the open-source LeRobot pipelines
- Bring the trained policy back to Raven
Skills Store
Install a skill in one tap.
Browse skills made by the community, from games to useful automations. Publish yours and it runs on every Raven.
- Skill Studio Record motions, write the AI prompt, design the control panel. Zero code.
- Games and automations Tic Tac Toe, Blackjack, a candy dispenser, a barista. Start with one, then make your own.
- Publish and share Publish a skill once and anyone with a Raven can install it.
Coming soon
Raven is almost here.
Join the waitlist for early-access pricing, a demo video, and the app preview before launch.