01 · Overview
See. Think. Move.
Lightweight modular Vision–Language–Action for robotics. Not one giant transformer: specialized modules talk through a World Model.
Pipeline
Instruction and perception feed a persistent world model; policy and control close the loop with feedback.
Instruction → Language → Planner → World Model
Vision / sensors → Scene → World Model
World Model + subtask → Policy → Control → Robot
↑____________ feedback ___________|Quick start
Requires Python 3.10+ (CI and primary dev: 3.11).
Public source for Kinetic is not on GitHub yet. Use the docs on this site for architecture and backends; CLI examples below describe the intended install path once the package is published.
# Once published: # pip install kinetic[dev,sim] # Hero path (60 seconds) kinetic demo pick # always works (mock) # or physics + 3D window: kinetic demo mujoco # needs MuJoCo + display # Interactive CLI kinetic
# More kinetic demo list kinetic run -i "Pick up the bottle" kinetic train-bc # demos → MLP → eval kinetic eval --seeds 5 kinetic vla-bench --trials 5 pytest -q
Optional backends via config: control.ik_backend: dls, scene.backend: yolo, policy.backend: lerobot (SmolVLA/ACT). See Backends.
What this is / isn't
This is
Modular VLA software stack (sim-first)
This is not
A single 7B end-to-end OpenVLA clone
This is
Event-driven language + world model
This is not
LLM joint commands every control tick
This is
Mock + MuJoCo demos you can run today
This is not
Turnkey industrial robot product
This is
Interfaces for ROS2 / real arms
This is not
Certified safety controller
This is
Latency + literature VLA comparison
This is not
LIBERO leaderboard without trained policies
This is
BC-trained sim policy + multi-task eval
This is not
A frontier foundation model (yet)
Support matrix
See the roadmap for the honest gaps list. Kinetic source is not public yet — docs on this site are the source of truth.