flex-pi/flexpi-libero-fulljoint-star

Flex-π · LIBERO — full joint, no 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. Actions are generated jointly with those futures.

This checkpoint was fine-tuned on LIBERO with every stream always present and always jointly denoised — the Flex-π* row of the paper. Evaluate it full joint.

Results

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

Inference regime	Success (%)
full joint	99.2

It never saw a sample with a missing stream, so the action-only path and missing-modality deployment are not its job. For those, flex-pi/flexpi-libero serves both regimes from one set of weights at 98.4 / 98.5.

Model details

Training data	flex-pi/libero_mujoco3.3.2_depth
Action space	32-D, rotvec, arm-grouped
Cameras	agentview + wrist, composited at 448×512 (tshape_libero_2cam_448x512)
Visual streams	RGB and 3D pointmap through the frozen Wan-2.2 VAE, DINOv3 ViT-B/16 folded 2×2
Flex regime	p_present_* = p_j* = 1.0 — no dropout
Checkpoint	step_021690, bf16, 12.1 GB

Usage

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

The six regime flags default to true, which is what this checkpoint wants:

CKPT=./checkpoints/flexpi-libero-fulljoint/checkpoints/weights/step_021690.pt \
DATASET_STATS=./checkpoints/flexpi-libero-fulljoint/dataset_stats.json \
GPUS=0,1,2,3,4,5,6,7 \
  bash scripts/eval_flexpi_libero_4suite.sh

Turning streams off here evaluates a path this model never trained on — the FlexPi inference regime: line echoed at startup should read denoise=(action=True, video=True, dino=True, pointmap=True). 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.