Haodong082399/VPP2

Video Prediction Policy 2: Predict Better, Act Better

Official VPP2 weights for RoboDojo and LIBERO.

Paper · Code · Project page · ModelScope

Model weights

Checkpoint	Path	Use
RoboDojo history-conditioned Video-10k	checkpoints/initialization/robodojo_his10k.pt	Joint + Action2B training initializer
RoboDojo joint Video-100k	checkpoints/joint2b_s100000/video.pt	Paired RoboDojo evaluation
RoboDojo Action2B-100k	checkpoints/joint2b_s100000/action.pt	Paired RoboDojo evaluation
LIBERO Video-10k	checkpoints/libero/video_step010000.pt	Action training and evaluation
LIBERO Action-30k	checkpoints/libero/action_step030000.pt	LIBERO, LIBERO-OOD and LIBERO-PRO evaluation

Shared VAE, CLIP, UMT5 and tokenizer assets are under checkpoints/Wan2.1-I2V-14B-480P/. Each policy pair includes its manifest and normalization statistics. Keep each pair together.

TODO

• Release the robot-video pretrained backbone checkpoint.

Download

python -m pip install -U huggingface_hub
hf download Haodong082399/VPP2 --local-dir weights

For only the RoboDojo evaluation bundle:

hf download Haodong082399/VPP2 --local-dir weights \
  --include 'checkpoints/joint2b_s100000/*' --include 'checkpoints/Wan2.1-I2V-14B-480P/*'

The directory structure matches the ModelScope release. Follow the official RoboDojo guide or LIBERO guide for installation, training and evaluation.

Evaluation settings

Benchmark	Checkpoint pair	Denoising	Action horizon / execution
RoboDojo	Video-100k + Action2B-100k	10 Euler steps, shift 1	32 / 24
LIBERO / OOD / PRO	Video-10k + Action-30k	10 Euler steps, shift 5	32 / 10

RoboDojo uses its EE16 z-score statistics; LIBERO uses its own 7D action min/max statistics. Use each benchmark's configuration and simulator setup.

Citation

See CITATION.bib for the paper citation.

Licenses

Shared Wan assets retain their upstream license in checkpoints/Wan2.1-I2V-14B-480P/LICENSE.txt. The official implementation has its own code license.