Inference orchestration

An orchestrator decides how a model interacts with motor execution: when to run inference, how to handle interruptions, and when to inject recovery actions. You attach one at construction, and the robot controller delegates every step to it.

from cadenza.inference import Sequential, ChainOfThought
import cadenza_lab as cadenza

go1 = cadenza.go1(inference=Sequential())
go1.run([go1.walk_forward(distance_m=5.0)])

Without an orchestrator, go1.run([...]) executes open-loop, with no model in the loop. With one, each Step is run under the strategy’s control.

`Sequential`: think then move

Single-threaded “guard while you act”. For each action, a guardian (default: VLAGuardian, a small VLM on the forward camera) watches for obstacles. On a detection mid-action:

The current step is cut short at whatever fraction completed.
The guardian emits an avoidance sequence (turn / side-step / wait).
Avoidance runs with the guardian off (no recursive interrupts).
The original action resumes with the remaining distance, bounded by retries so the robot can’t thrash.

Sequential(*, show_camera=True, model_id=None, min_resume_distance_m=0.1,
           retries=5, guardian=None, logging=None)

Param	Description
`show_camera`	Show the guardian’s live camera window.
`model_id`	Override the guardian’s VLM checkpoint.
`min_resume_distance_m`	If less than this remains after an interrupt, drop the action instead of resuming.
`retries`	Max VLA interrupts per step before abandoning (`None` = ∞).
`guardian`	Custom detector: a class, factory, or instance.
`logging`	Path to a JSON-Lines event log (for later training).

`ChainOfThought`: think while you move

Concurrent inference and execution. While the robot executes action N, a background worker is already running the next inference pass to produce N+1, so there’s no dead time waiting on the model.

tick 0  main: ───── exec A0 ─────
        bg:             [ infer A1 ]
tick 1  main: ───── exec A1 ─────
        bg:             [ infer A2 ]

ChainOfThought(*, model=..., sense=[...], goal="...", target=(x, y),
               max_steps=80, logging=None)

The model is any WorldModelAdapter. sense is a list of modalities merged into each observation. Any trigger Step you pass to run() hands control to the orchestrator.

Demo: guarded walk

"""demo_orchestration.py: guarded walk with the Sequential strategy.
Opens a MuJoCo window (needs a display). The guardian needs the `ai` extra
(torch) and downloads a small VLM on first run.   pip install "cadenza-lab[ai]"
"""
import cadenza_lab as cadenza
from cadenza.inference import Sequential

go1 = cadenza.go1(
    xml_path=cadenza.Go1.terrain("terrain"),
    inference=Sequential(show_camera=True, retries=5, logging="runs/episode.jsonl"),
)
go1.run([go1.walk_forward(distance_m=5.0)])   # guardian interjects around obstacles

pip install "cadenza-lab[ai]"
python demo_orchestration.py

Orchestration controls timing and interruption. To swap the actual decision model, or to run a goal-directed loop end to end, use the inference stack.

Get started

Features

Tutorial

Inference orchestration

`Sequential`: think then move

`ChainOfThought`: think while you move

Demo: guarded walk

​Sequential: think then move

​ChainOfThought: think while you move

​Demo: guarded walk

`Sequential`: think then move

`ChainOfThought`: think while you move

Demo: guarded walk