Robot Actions

This is why strands-diffusers exists. A world-foundation model like NVIDIA Cosmos doesn't just generate a plausible future video - it predicts the robot action chunk that produces it. use_diffusers serializes that chunk and can render it so you can see the motion.

A reach-grasp-lift, visualized straight from the action tensor.

See an action chunk

from strands_diffusers import use_diffusers

# action shape: [num_chunks, T, action_dim] - a list, .json path, or cached:key
use_diffusers(action="visualize", inputs=action_chunk,
              parameters={"fps": 8})

This produces three artifacts:

time-series (every dim over time, gripper highlighted)	end-effector path (dims 0-2 integrated)

...plus an .mp4 animation sweeping a playhead across the curves - optionally side-by-side with the generated world video frames.

The action payload

A Cosmos Cosmos3OmniPipelineOutput carries video, optional sound, and action. The serializer writes each to the right artifact:

{
  "type": "action",
  "chunk_shape": [16, 10],
  "num_chunks": 1,
  "path": "/tmp/strands_diffusers/action_*.json"
}

The .json is the model-normalized action chunk (values in [-1, 1]) - feed it straight to your embodiment's un-normalizer and controller. Values are preserved exactly (no lossy clipping); bf16 tensors are upcast to f32 before serialization.

End to end

See World-Foundation Models for the full Cosmos action-policy rollout that produces video and actions from one use_diffusers(action="run", ...) call.