IMLE-VLA checkpoints (IROS 2026)

Checkpoints for IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies (paper, project page, code).

Five-minute teaser: the setting, why regression collapses, why flow matching is slow and jerky, IMLE-VLA on the real robot, and how conditional IMLE works.

IMLE-VLA replaces the 10-step flow-matching action head of π₀.₅ with a single-step generator trained with conditional IMLE. Same frozen VLM backbone, one forward pass per action chunk, 55 Hz instead of 15 Hz on an NVIDIA L40S.

Contents

path	what	size
droid/jax/57000/	JAX (orbax) checkpoint of the real-robot DROID model: params/ + assets/droid/norm_stats.json. Training config 16_horrizon_05_lr_2_factor_96_batch, step 57000. Serve with scripts/serve_policy.py policy:checkpoint.	6.6 GB
droid/triton/	The same weights converted for the single-step Triton engine (pi05_imle_droid_57000.pkl) plus assets/droid/norm_stats.json, the PaliGemma tokenizer and a provenance copy of the engine. This is the 55 Hz path used on the robot.	6.3 GB

The DROID model: π₀.₅ fine-tuned on DROID with the cIMLE objective (sample factor m = 2), frozen VLM, trained action expert only, joint-position action space, 16-step action chunks, two camera views (one exterior + wrist). It is the policy behind the real-robot experiments in the paper (Franka Emika Panda, four tabletop tasks, 55 Hz at execution horizon 12).

Real-robot videos

Franka Emika Panda, both policies trained on DROID: π₀.₅ at 15 Hz versus IMLE-VLA at 55 Hz (the droid/ checkpoint).

Reactive task: pineapple on a moving plate. A remote-controlled car moves the plate, so the target is in continuous motion.

IMLE-VLA	π₀.₅

Static tasks (side by side: π₀.₅ left, IMLE-VLA right)

Swap pineapple and cube

Pineapple in bowl

Pineapple in cabinet

Use

git clone --recurse-submodules https://github.com/kianhk6/IMLE-VLA.git && cd IMLE-VLA/openpi   # install per openpi/README.md
pip install -U huggingface_hub

# JAX checkpoint
huggingface-cli download kinster/IMLE-VLA --include "droid/jax/*" --local-dir ./ckpt
python scripts/serve_policy.py --env=DROID --port=8000 policy:checkpoint \
    --policy.config=16_horrizon_05_lr_2_factor_96_batch \
    --policy.dir=./ckpt/droid/jax/57000 \
    --policy.sample-kwargs='{"num_steps": 1}'

# Triton package (55 Hz)
huggingface-cli download kinster/IMLE-VLA --include "droid/triton/*" --local-dir ./ckpt
scripts/serve_triton.sh ./ckpt/droid/triton 16_horrizon_05_lr_2_factor_96_batch 0 8000 DROID

Then run the DROID robot client (examples/droid/main.py --open_loop_horizon 12) or score the served policy against ground-truth DROID actions with scripts/validate_triton_droid.py. Both are documented in openpi/README.md and openpi/docs/triton.md.

LIBERO

libero/imle_vla_libero_h10_{spatial,object,goal,long,overall} and libero/imle_vla_libero_h30_{...} hold LeRobot checkpoints (pretrained_model/), per suite and overall, for execution horizon 10 and 30. Train and evaluate them with the LeRobot code in the repository's lerobot/ folder.

Citation

@inproceedings{Hosseinkhani2026IMLEVLAFS,
  title     = {{IMLE-VLA}: Fast Single-Step Action Generation for Vision-Language-Action Policies},
  author    = {Hosseinkhani, Kian and Peng, Qinhe and Shramko, George and Aghabozorgi, Mehran and
               Qian, Jianing and Engst, Tristan and Moazeni, Alireza and Jayaraman, Dinesh and Li, Ke},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  eprint    = {2609.10915},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2609.10915}
}

License

Apache 2.0, like the openpi code and π₀.₅ checkpoints these weights are fine-tuned from. The PaliGemma backbone is subject to the Gemma Terms of Use.