1.3.2

NOW COMBINE YOUR OWN MODELS to build effective physical AI stacks ON CADENZA!

Multi-modal sensing added to help your inserted VLA model process data (e.g. depth estimation) and make effective actions.

Along so, a NEW ACTION LOOP is integrated, allowing your VLA model to communicate in Cadenza's action library and is proactive, allowing for easier debugging and improve RL systems.

import cadenza_lab as cadenza
import cadenza_lab.stack
from depth_anything_v2_small import DepthAnythingV2Small
from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

cadenza.stack.register_world_model(
"smolvla", model=SmolVLAPolicy.from_pretrained("lerobot/smolvla_base"),
)

cadenza.stack.run(
robot="go1",
goal="reach the green beacon at the top of the stairs and sit",
target=(-5.5, 0.0),
xml_path="examples/smolvla-test/stairs_scene.xml",
modalities=[DepthAnythingV2Small()],
)

Full Changelog: https://github.com/aparekh02/cadenza/commits/latest