Flex-π — RoboTwin 2.0 (3-camera, 384×320)
Flex-π checkpoint for RoboTwin 2.0, trained jointly on all 50 tasks.
- Paper: arXiv:2608.10860
- Project page: https://flex-pi.github.io/
- Code: https://github.com/geyan21/flex-pi
Results
RoboTwin 2.0, success rate (%) over 50 tasks, as reported in the paper (Table 1). Clean and Randomized are background conditions; the two rows are inference regimes served by these same weights.
| Inference regime | Clean | Randomized | Avg. |
|---|---|---|---|
| Action-only | 94.5 | 94.6 | 94.6 |
| Full joint | 94.3 | 94.8 | 94.6 |
K = 4 Euler denoising steps throughout. Full joint additionally denoises the
future-video, DINO, and pointmap streams; action-only skips them, trading them
for cheaper inference at the same average success rate. Selecting a regime is an
inference-time flag — no retraining, no separate weights.
Architecture
A Mixture-of-Transformers pairing a video DiT with an action DiT, coupled by HBridge. Alongside actions the model can denoise three auxiliary streams: future video, DINO features, and pointmaps.
| Video expert | Wan2.2-TI2V-5B, 5.00 B params |
| Action expert | ActionDiT, 1.02 B params |
| Layers | 30 (HBridge: 7 bottom / 16 middle / 7 top) |
| Semantic encoder | DINOv3 vit_base_patch16_dinov3.lvd1689m, 768-d, frozen |
| Cameras | cam_high, cam_left_wrist, cam_right_wrist @ 240×320 |
| Composite video | 384×320, 33 frames |
| Action / proprio | 14-d each (bimanual ALOHA-AgileX), ConcatLeftAlign |
| Action : video rate | 4:1 |
Training
| Data | 2,500 clean + 25,000 randomized demos, all 50 tasks |
| Epochs | 6 |
| Learning rate | 1e-4 |
| Precision | bf16 |
Trained with flex-joint sampling at p = 0.5 on every present and joint flag,
with cross-modal prediction enabled for all three streams. That is what lets one
set of weights serve any regime in the results table above.
Files
config.yaml # architecture + processor; autoloaded by the eval
dataset_stats.json # action/state normalization statistics
checkpoints/weights/step_048060.pt # 12 GB
Keep this directory layout. The eval locates config.yaml and
dataset_stats.json by walking up from the checkpoint path.
Usage
This repository holds the policy weights only. The Wan2.2 base components
and the ActionDiT backbone are separate downloads, resolved through
DIFFSYNTH_MODEL_BASE_PATH — see docs/INSTALL.md
and docs/ROBOTWIN.md.
hf download flex-pi/flexpi-robotwin --local-dir runs/flexpi-robotwin
export DIFFSYNTH_MODEL_BASE_PATH="$(pwd)/checkpoints" # Wan2.2 weights
Then set the checkpoint at the top of scripts/eval_flexpi_robotwin.sh:
CKPT="./runs/flexpi-robotwin/checkpoints/weights/step_048060.pt"
DATASET_STATS="./runs/flexpi-robotwin/dataset_stats.json"
and run:
bash scripts/eval_flexpi_robotwin.sh
The launcher defaults reproduce the full joint row: NUM_INFERENCE_STEPS=4,
INSTRUCTION_TYPE=unseen, EVAL_NUM_EPISODES=100, and all six regime flags
true. For the action-only row, set the three INFER_JOINT_* flags to
false. PHASES=clean,random covers both background conditions.
License
MIT — see LICENSE.
Citation
@article{yan2026flexpi,
title = {Flex-$\pi$: A Multi-Stream World-Action Model with Compute Flexibility},
author = {Yan, Ge and Liu, Jinghao and Fan, Yuzhi and Cai, Lei and Liao, Minwen
and Zhang, Jesse and Fox, Dieter},
journal = {arXiv preprint arXiv:2608.10860},
year = {2026},
url = {https://arxiv.org/abs/2608.10860}
}