Perceive
Vision and sensors feed scene understanding into the World Model. Optional YOLO backend for pretrained detection.
A MODULAR VLA STACK
robots that see, think move
Lightweight modular Vision–Language–Action for robotics. Not one giant transformer, specialized modules talk through a World Model.
SIM-FIRST · MOCK + MUJOCOEVENT-DRIVEN WORLD MODELAPACHE-2.0 OPEN SOURCE↓ 60-SECOND DEMO
Vision–Language–Action at the Intersection ofSoftware & Robotics
World Model is the integration bus. Language never emits joint commands. Safety lives in the motion controller. LLM and planner run on events, not every tick.
THE PIPELINE
How it moves
Understand
Language and planner turn instructions into subtasks, event-driven, not every control tick. Optional BitNet for local reasoning.
Act
Policy → control → robot. BC-trained policy, damped-least-squares IK, and optional LeRobot/SmolVLA pretrained checkpoints.
Learn
train-bc for behavior cloning, multi-task eval across pick and place, and vla-bench for systems benchmarks.
A Stack You CanRun Today
| Backend | Detail | Entry | Status |
|---|---|---|---|
| Mock sim | Always works, lightweight arm, no extra deps | kinetic demo pick | 60 s |
| MuJoCo | Physics + 3D viewer tabletop scene | kinetic demo mujoco | Optional |
| BitNet LLM | 1.58-bit local reasoner on replan events | configs/bitnet.yaml | Optional |
| LeRobot / SmolVLA | Pretrained HF VLA checkpoints through the loop | policy.backend: lerobot | ~27 s / 50-action chunk on CPU |
| ROS2 | Joint trajectory adapter for real arms | robots.ros2 | Dry-run |
See. Think.Move.
Benchmarks
