The stack at a glance
What’s included
Simulation core
A MuJoCo-based simulator with quadruped and humanoid robots (Go1, G1),
composable scenes, and contact-rich physics out of the box.
Action library
Reusable, phase-aware robot actions — gaits, manipulations, sequences — that
you call by name or compose into multi-phase missions.
Mission specs
Describe a multi-phase task as a single
env.json: the scene, the phases,
success/fail predicates, and reward shaping.Rollout logging
Every tick of every run is captured and scored by an LLM judge, giving you
structured, labeled trajectories automatically.
Fine-tuning pipeline
Convert rollouts into VLA training data and train LoRA adapters — RL and
supervised fine-tuning behind one command.
Inference stack
Orchestrate policies and world models — sequential, chain-of-thought, or your
own VLA — through a clean inference interface.
Deploy drivers
Push a tested policy to physical hardware over DDS, SSH, or a bridge, using
the exact loop you ran in simulation.
Natural-language control
Parse commands like “walk forward then sit” into action calls — a fast path
from intent to motion.
Who it’s for
- Robotics researchers prototyping policies and gathering training data without building a sim pipeline from scratch.
- Applied ML teams fine-tuning VLA models on task-specific rollouts.
- Robot product teams that need a reproducible path from simulation to a deployed unit.
Why Cadenza?
How our stack compares to the status quo.