lozoRUST/motus-value-step7500

Motus Value Model — Test Checkpoint (step 7500)

Value (progress-scoring) model test checkpoint finetuned from Motus (Stage 2 pretrained, Apache-2.0). Given a robot manipulation trajectory (multi-camera videos + task text), it outputs window-level value scores, text-match probability, and an overlap-averaged per-frame value curve.

Inference settings are fixed: 5 FPS sampling, 5-frame windows, stride 2. Verified on a single 48 GB RTX 4090 D (peak ≈ 23.1 GiB with batch size 16); reduce --batch-size to 4/8 for smaller GPUs.

Files

File	Size	SHA-256
checkpoints/step_7500.pt	3,564,872,186	cb8005fb57ccb8627553dde5e25dfe22150d47b14b49de017954b165f859acd7

This checkpoint contains the 7500-step trainable parameters and does not work standalone — the base weights below are required.

Required base weights (NOT included — download yourself)

Place everything under a single pretrained_models/ directory next to checkpoints/:

Component	Source	~Size	Layout
Motus (Stage 2 pretrained)	thu-ml/Motus	~16 GB	pretrained_models/Motus/mp_rank_00_model_states.pt
Wan2.2 VAE + UMT5-xxl encoder	Wan-AI/Wan2.2-TI2V-5B	~14 GB	pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth, pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth, pretrained_models/Wan2.2-TI2V-5B/google/umt5-xxl/
Qwen3-VL-2B-Instruct tokenizer / processor configs	Qwen/Qwen3-VL-2B-Instruct	~12 MB	pretrained_models/Qwen3-VL-2B-Instruct/

Reference checksums of the exact files used during validation (SHA-256):

checkpoints/step_7500.pt                                cb8005fb57ccb8627553dde5e25dfe22150d47b14b49de017954b165f859acd7
pretrained_models/Motus/mp_rank_00_model_states.pt      a9b643f4cf087c543d3741a8ecbd9a158e6f3112c5d0f3e53f19ed16f682779a
pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth  7cace0da2b446bbbbc57d031ab6cf163a3d59b366da94e5afe36745b746fd81d
pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth         20eb789667fa5e60e7516bf509512f6cb61f01b0aa0695eadaea930c13892b36
pretrained_models/Wan2.2-TI2V-5B/google/umt5-xxl/tokenizer.json   6e197b4d3dbd71da14b4eb255f4fa91c9c1f2068b20a2de2472967ca3d22602b

Usage

Input contract (per trajectory): head-camera video + left/right wrist videos + task text (e.g. "push tray to customer"). The model scores each 5-frame window and averages overlapping windows into a per-frame value curve; higher value ≈ more on-task progress.

Command-line sketch (see the training/inference code this checkpoint was validated with):

run_inference.sh \
  --head /path/head.mp4 --left-wrist /path/left.mp4 --right-wrist /path/right.mp4 \
  --text "push tray to customer" \
  --checkpoint checkpoints/step_7500.pt \
  --output outputs/result.json

License

• This checkpoint is a derivative of Motus (Apache-2.0) and is released under Apache-2.0.
• Base/third-party weights remain under their own licenses (Motus: Apache-2.0 · Wan2.2: Apache-2.0 · Qwen3-VL: Apache-2.0).

Citation

@misc{motus2025,
  title = {Motus: A Unified Latent Action World Model},
  howpublished = {\url{https://github.com/thu-ml/Motus}},
  note = {arXiv:2512.13030}
}