Cerebellum: Vision-Language-Action Skills

SUMMARY

Cerebellum is a module in the Telekinesis Agentic OS for Vision-Language-Action (VLA) and World Action Model (WAM) skills in robotics: it fine-tunes imitation-learned manipulation policies - ACT, π0.5, and Fast-WAM - on your own demonstrations, then deploys them for closed-loop robot control from camera and proprioceptive input. A policy is described by a single YAML file, so fine-tuning a different task means pointing at a different config and dataset.

Install Telekinesis Skill Library

Create an API key, install telekinesis-ai and module specific installations and dependencies in a Python environment.

Open installation →

Go to the Quickstart

Pick a starter and run your first Skill end-to-end - 2D / 3D vision, webcam capture, robot motion, or a full vision-to-robot pipeline, all visualized in Rerun.

Open quickstart →

When to Use Cerebellum?

Use Cerebellum for manipulation tasks that are easier to show than to program, such as:

• Fine-tuning ACT, π0.5, or Fast-WAM on teleoperated demonstrations for a specific pick, insertion, or bimanual task
• Evaluating a policy against held-out rollouts before it ever touches hardware
• Deploying a policy for closed-loop control, mapping live camera and joint-state input to actions at control-loop rate
• Integrating learned manipulation behavior into Physical AI pipelines alongside perception, planning, and classical control

What Does Cerebellum Provide?

Cerebellum includes a collection of modular skills for:

• Formatting teleoperated demonstrations into a training dataset
• Fine-tuning ACT, π0.5, and Fast-WAM policies from a single YAML configuration
• Evaluating a fine-tuned policy against recorded or live rollouts
• Deploying a policy for closed-loop inference on a real manipulator through one interface, regardless of which VLA or WAM backend produced it

Guides

Fine-tune a VLA policy on your own demonstrations, evaluate it against held-out rollouts, and run it in closed loop on the robot.

ACT

Fine-tune and deploy an Action Chunking Transformer policy - sample-efficient imitation learning for fine, bimanual manipulation.

Start here →

π0.5

Fine-tune and deploy Physical Intelligence's π0.5 - a pretrained, language-conditioned generalist policy for open-world generalization.

Explore →

Fast-WAM

Fine-tune and deploy Fast-WAM - a World Action Model that trains on video-predicted futures but skips them at test time for real-time control.

Explore →