yygx/lilo-vla-openvla-oft

LiLo-VLA — interaction policy (OpenVLA-OFT, LoRA r32)

The atomic-skill policy used by LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies. It is the Interaction Module M_int: given a wrist-camera image, object-relative proprioception and a skill instruction, it emits an 8-step action chunk.

Code: https://github.com/YY-GX/LiLo-VLA

What this is

Backbone	OpenVLA-OFT 7B, LoRA rank 32
Action head	L1 regression, 8-step chunks, 7-DoF
Observation	wrist camera only (224x224, centre-cropped), 8-D proprio
Training	200k steps, batch 16, lr 5e-4, --image_aug
Training data	yygx/lilo-vla-atomic-skills-rlds

It is not a standalone long-horizon policy. On its own it performs one atomic skill from an approach pose. The reaching module, skill verifier and recovery loop live in the code repository.

Usage

pip install -e ".[libero,method]"     # from the LiLo-VLA repo

python scripts/evaluate_long_horizon.py \
    --vla_checkpoint yygx/lilo-vla-openvla-oft \
    --task_name 'Complete Kitchen Organization (Original)' \
    --num_trials 10 \
    --benchmark ultra_long \
    --apply_distractor_masking

dataset_statistics.json ships with the weights and is what selects the action de-normalisation statistics, so the checkpoint works from a Hub id without extra flags.

Evaluation

Measured with this checkpoint on the 21-configuration LiLo-VLA benchmark, 10 trials per configuration. Success Rate = the entire skill sequence completed in order; Average Progress = mean fraction of the sequence completed.

Suite	SR	AP
LIBERO-Long++ (12 configurations)	80.0%	89.6%
Ultra-Long (9 configurations)	see repo	see repo

The benchmark's evaluation is unseeded, so per-task SR varies run to run; Average Progress is the stable metric.

License

MIT. Built on OpenVLA-OFT (MIT, (c) Moo Jin Kim, Chelsea Finn, Percy Liang).

Citation

@article{yang2026lilo,
  title={LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies},
  author={Yang, Yue and Cheng, Shuo and Fang, Yu and Bharadhwaj, Homanga and Ding, Mingyu and Bertasius, Gedas and Szafir, Daniel},
  journal={arXiv preprint arXiv:2602.21531},
  year={2026}
}