Flex-π · LIBERO — stream dropout

Flex-π is a 6B-parameter world-action model that supervises three visual futures — RGB, 3D pointmaps and DINOv3 semantics — from the RGB you already have. A single checkpoint runs on any subset of its streams, so the compute/accuracy operating point is a runtime flag.

This checkpoint was fine-tuned on LIBERO with every stream probability at 0.5, so it serves both the fast action-only path and full joint generation.

Results

LIBERO, 4 suites × 10 tasks × 50 trials, official per-suite step budgets (220 / 280 / 300 / 520), mujoco 3.3.2.

Both from this one checkpoint. Trained without dropout instead, flex-pi/flexpi-libero-fulljoint-star scores 99.2 full joint and gives up the action-only mode.

Model details

Usage

huggingface-cli download flex-pi/flexpi-libero --local-dir ./checkpoints/flexpi-libero

CKPT=./checkpoints/flexpi-libero/checkpoints/weights/step_010860.pt \
DATASET_STATS=./checkpoints/flexpi-libero/dataset_stats.json \
GPUS=0,1,2,3,4,5,6,7 \
  bash scripts/eval_flexpi_libero_4suite.sh

The action-only path is the same command plus three flags:

INFER_JOINT_VIDEO=false INFER_JOINT_DINO=false INFER_JOINT_POINTMAP=false \
CKPT=... DATASET_STATS=... bash scripts/eval_flexpi_libero_4suite.sh

The regime is not inferred from the checkpoint, and a wrong choice produces a plausible but meaningless number with no error — check the FlexPi inference regime: line echoed at startup. Evaluation needs ~15 GB of VRAM. config.yaml and dataset_stats.json have to stay beside the weights: the architecture is read back from the saved config, never respecified on the command line.

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}
}

MIT licensed.