cadenza-lab 1.6.0

Cadenza Lab - The SDK to build in physical AI

Run and deploy complex robot actions with a simple Python SDK.

Features • Quickstart • Projects • CLI • Deploy

Cadenza lets you simply write complex motion-targeted code and deploy on MuJoCo or hardware for Unitree Go1 (quadruped) and G1 (humanoid) robots.

Website: www.cadenzalabs.xyz  •  pip package: cadenza-lab

Want to integrate your own hardware? Cadenza targets the Go1 and G1 today. If you'd like support for drones, robot arms, or any other robot, reach out — [e-mail].

⭐ Features

Action Library

• 41 motor-level primitives across Go1 (21 actions) and G1 (20 actions) — joint targets, PD gains, and torque limits sourced directly from URDF
• Phase-based actions: stand, sit, lie down, stand up, jump, rear up, shake hand
• Gait-based actions: walk, trot, crawl, pace, bound, climb, turn, sidestep
• Composable: run actions sequentially or concurrently in a single call
• Parameterized: every action accepts speed, extension, distance_m, and repeat

Simulation

• MuJoCo simulator built in — test any action sequence before touching hardware
• Natural language commands: cadenza sim go1 "walk forward then jump" just works
• World-model loop: attach a VLA model + perception modalities with setup(), hand the robot a goal, and Cadenza closes the sense → think → act loop
• Bring your own models: concrete VLA / depth / RGB models live in the ai_models/ package — swap in your own without touching the framework

Deploy

• SSH deploy: upload and run a script on the robot's onboard computer
• DDS direct: send motor commands from your laptop over DDS (same network)
• Bridge mode: run heavy compute on your laptop, lightweight actions on the robot

⚡ Quickstart

pip install cadenza-lab

Scripted actions

Compose primitives into a sequence and run it in MuJoCo — the Go1 stands, walks 2m, arcs through a turn, jumps, and sits. Nest actions in a list to run them concurrently.

import cadenza

go1 = cadenza.go1()
go1.run([
    go1.stand(),
    go1.walk_forward(speed=1.5, distance_m=2.0),
    [go1.turn_left(), go1.walk_forward()],   # concurrent: walking arc
    go1.jump(speed=2.0, extension=1.2),
    go1.sit(),
])

World-model driven

Attach a VLA model and perception modalities with setup(), then hand the robot a natural-language goal. Cadenza closes the loop: sense → think → act.

import cadenza
from ai_models.go1 import VLA, Depth, RGB

go1 = cadenza.go1()
go1.setup(
    model=VLA(),
    sense=[Depth(), RGB()],
)
go1.run(
    goal="reach the green beacon at the top of the stairs and sit",
    scene="stairs",
    target=(-5.5, 0.0),
)

Cadenza ships the framework — base adapters, perception modalities, and the runtime. The concrete models live in the ai_models/ package alongside your project, so you can swap in your own VLA or perception stack. The repo's example.py runs this demo end-to-end: mjpython example.py.

G1 Humanoid

import cadenza

g1 = cadenza.g1()
g1.run([
    g1.stand(),
    g1.walk_forward(distance_m=1.0),
    g1.crouch(),
    g1.hold(duration=1.0),
    g1.stand(),
])

🤖 Free Projects

Community projects built with Cadenza. Add yours via a pull request.

Project	Robot	Description	Link
Go1 Obstacle Course	Go1	VLA-guided navigation through a MuJoCo obstacle course	coming soon
G1 Gesture Control	G1	Map hand gestures to G1 arm actions via webcam	coming soon
Multi-robot Sync	Go1 + G1	Synchronized action sequences across two robots	coming soon

🖥️ CLI

cadenza list go1                                      # list all Go1 actions
cadenza list g1                                       # list all G1 actions
cadenza sim go1 "walk forward then jump"              # simulate in MuJoCo
cadenza sim g1 "stand then walk forward"              # simulate G1
cadenza sim go1 "walk forward" --vla --obstacles      # VLA obstacle avoidance
cadenza deploy go1 --ip 192.168.123.15 -c "..."       # deploy via SSH
cadenza deploy go1 --ip ... --mode direct             # deploy via DDS
cadenza deploy go1 --ip ... --mode bridge             # bridge mode

📦 Action Library Reference

Go1 — 21 actions

Action	Type	Description
stand()	phase	Stand at default height
stand_up()	phase	Stand up from lying down
sit()	phase	Sit down
lie_down()	phase	Lie flat
jump()	phase	Jump in place
rear_up()	phase	Rear up on hind legs
shake_hand()	phase	Extend front paw
rear_kick()	phase	Kick with rear legs
walk_forward()	gait	Walk forward
walk_backward()	gait	Walk backward
trot_forward()	gait	Trot (diagonal gait)
crawl_forward()	gait	Crawl (low, stable)
pace_forward()	gait	Pace (lateral gait)
bound_forward()	gait	Bound (synchronous front-back)
turn_left()	gait	Turn left in place
turn_right()	gait	Turn right in place
climb_step()	gait	Climb a step
side_step_left()	gait	Lateral step left
side_step_right()	gait	Lateral step right

All actions accept speed and extension multipliers. Gait actions also accept distance_m and repeat.

G1 — 20 actions

Access via cadenza.g1(). Full action list: cadenza list g1.

from cadenza.actions import get_library, list_actions

list_actions("go1")               # print all actions

lib = get_library("go1")
spec = lib.get("walk_forward")
print(spec.gait)                  # GaitAction with velocity commands

🐾 Go1 Quadruped

The Go1 is a quadruped robot with 12 joints across four legs. Cadenza provides 21 actions for it.

mjpython example.py                                  # run the Go1 demo
cadenza sim go1 "walk forward then jump"             # simulate via CLI
cadenza list go1                                     # list all Go1 actions
cadenza deploy go1 --ip 192.168.123.15 -c "..."      # deploy to hardware

import cadenza

go1 = cadenza.go1()
go1.run([
    go1.stand(),
    go1.walk_forward(speed=1.5, distance_m=2.0),
    go1.jump(),
    go1.sit(),
])

🤖 G1 Humanoid

The G1 is a full-size humanoid robot. Cadenza provides 20 actions for bipedal locomotion and arm control.

cadenza sim g1 "stand then walk forward"             # simulate via CLI
cadenza list g1                                      # list all G1 actions
python examples/unitree_g1/deploy_g1.py sim          # run the G1 example

import cadenza

g1 = cadenza.g1()
g1.run([
    g1.stand(),
    g1.walk_forward(distance_m=1.0),
    g1.crouch(),
    g1.hold(duration=1.0),
    g1.stand(),
])

Or hand it a goal with a world model:

import cadenza
from ai_models.g1 import VLA, Depth, RGB

g1 = cadenza.g1()
g1.setup(model=VLA(), sense=[Depth(), RGB()])
g1.run(goal="walk to the chair and sit", target=(2.0, 0.0))

🚀 Deploy

Run the same actions on real hardware. Three modes:

Mode	What it does	Command
SSH	Upload and run a script on the robot's onboard computer	cadenza deploy go1 --ip 192.168.123.15 -c "walk forward then sit"
DDS direct	Send motor commands from your laptop over DDS (same network)	cadenza deploy go1 --ip 192.168.123.15 --mode direct -c "..."
Bridge	Heavy compute on your laptop, lightweight actions on the robot	see below

Bridge mode keeps a live control handle for model-in-the-loop control:

go1 = cadenza.go1()
bridge = go1.deploy_ssh_bridge(host="192.168.123.15", key="~/.ssh/go1_rsa")

while True:
    action = my_model(bridge.telemetry)
    bridge.send_action(action, speed=0.5)

bridge.estop()

💚 Community

	Links
GitHub	aparekh02/cadenza
Issues	Report a bug or request a feature
License	Apache 2.0