---

name: model-inference description: "Run or plan Motus CUDA inference for real-world images and RoboTwin policy evaluation, including checkpoint loading, T5 conditioning, output contracts, and VRAM troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Motus model inference

Use this route for Motus image-to-video/action prediction, pre-encoded T5 conditioning, checkpoint compatibility, or RoboTwin policy deployment.

• Read api-reference.md for model/checkpoint contracts and tensor shapes.
• Read workflows.md for real-world and RoboTwin command recipes and asset preflight.
• Read troubleshooting.md for VRAM, CUDA, flash-attn, checkpoint, image, and external-runtime failures.
• Use data-preparation for dataset/image layout preparation and training for training configuration.

Operating procedure

1. Select real-world no-environment inference or RoboTwin evaluation. Confirm the checkpoint family, embodiment YAML, WAN/VAE assets, VLM checkpoint, and input image format before loading a model.
2. For real-world inference, provide one three-view concatenated RGB image, state/config matching the embodiment, and a language instruction. Prefer a pre-encoded WAN T5 tensor when VRAM is limited.
3. Run the parser/help preflight, then verify CUDA, GPU memory, file paths, and action/image dimensions. Actual model construction and denoising are CUDA-only and require large downloaded weights; do not substitute a CPU import for an inference result.
4. Inspect both returned artifacts: predicted future frames are [B,C,F,H,W] in [0,1], and actions are [B, action_chunk_size, action_dim].
5. For RoboTwin, configure the external simulator/runtime and policy paths separately. The policy route may execute robot/simulation side effects and is not a safe generic smoke test.

Launches, downloads, and simulator runs require explicit user approval.