Description

New Features: two new IL policies (VITA, FM + DiT)
Bug Fixes

Type of change

1. VITA Policy [0]

VITA (Vision-to-Action flow matching)[0] is a new fast and performant policy learning algorithm. VITA directly flows from latent images to latent actions without sampling from Gaussian or injecting conditions during denoising.

• The PR provides a VITA implementation with all the losses and network components used in the paper.
• Add flow matcher wrappers. Includes conditional FM, Optimal Transport FM [3], Schrodinger Bridge FM [4], etc., such that users can swtich to different FM algorithms via modifying the configs. The flow matcher wrapper provide an Euler solver; loss computation and sampling functions that supports custom source distributions.

2. Flow Matching Policy

RoboVerse now supports FM + UNet. The PR provides a DiT implementation that has been well tested in our previous papers and can outperform FM + UNet and match RoboVerse DP implementation.

• Add Diffusion Transformer [1] (DiT, as in Vision Transformer) using AdaLN blocks
• Add RoPE [2] to positional embeders for DiT

Local performance tests on CloseBox L0. We use the default collect_demos scripts to generate datasets and default DP runner configurations for all the models and conduct two random runs locally. FM + DiT outperforms UNet and VITA outperforms both DP and FM.

[0] Gao, Dechen, et al. "VITA: Vision-to-Action Flow Matching Policy." arXiv preprint arXiv:2507.13231 (2025).
[1] Peebles, William, and Saining Xie. "Scalable diffusion models with transformers." Proceedings of the IEEE/CVF international conference on computer vision. 2023.
[2] Su, Jianlin, et al. "Roformer: Enhanced transformer with rotary position embedding." Neurocomputing 568 (2024): 127063.
[3] Tong, Alexander, et al. "Improving and generalizing flow-based generative models with minibatch optimal transport." arXiv preprint arXiv:2302.00482 (2023).
[4] Tong, Alexander, et al. "Simulation-free schr" odinger bridges via score and flow matching." arXiv preprint arXiv:2307.03672 (2023).

3. Minor fixes.

• Remove repeated BaseImagePolicy declartions in DP policies and use the base class within DP utils.
• roboverse_learn/il/data2zarr_dp.py can be interrupted when meta json is not successfully generated. Added try-catch.
• All lower-cased consistent naming for file names. DDPM_model -> ddpm_model.
• Fix a Posix path type error ckpt_name = args.checkpoint_path.split("/")[-1] + "_" + time_str in dp_runner.py is not compatible with newer pathlib versions.

How to test

Please describe how to test the change if applicable.

• VITA Policy:
  Set algo_choose to vita_model (dp_run.sh)

• FM Policy + DiT:
  Set algo_choose to fm_dit_model (dp_run.sh)

Local performance tests on CloseBox L0:

Screenshots / Videos

Please attach before and after screenshots or videos of the change if applicable.

Checklist

• I have run the pre-commit checks with pre-commit run --color=always --all-files
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have added my name to the CONTRIBUTORS.md or my name already exists there