glowingspirits/let-it-be-simple

Let It Be Simple — LIBERO checkpoints

Policy checkpoints for the paper Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models (arXiv:2606.05737). Training / evaluation code and the matching run configs live in github.com/bondtesty/let-it-be-simple.

Both are Flow-Matching π0-style policies trained on the LIBERO demonstrations. One-step decoding means a single forward pass from the pure-noise endpoint (x_t = x_0 = ε); the same weights also run with multi-step Euler integration.

Checkpoint	Variant	What it is	Params
pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4	Tiny	action horizon 10, training timestep shift α = 4	240.4 M
pi0_libero_simvla_k4_nop_h10_cosine_150k_bs256	Full encoder	action horizon 10, α = 3, SigLIP all 27 layers, Gemma-2B backbone	1414.3 M

Each directory follows the openpi layout (params/ + assets/ with the LIBERO normalisation stats), so it loads without recomputing anything.

Results (from the paper)

Tiny, α = 4, one decoding step — LIBERO success rate % (Table 1):

Spatial	Object	Goal	Long
96.4	99.6	96.8	85.2

Full encoder, α = 3 (Table 2) — one step vs 10-step Euler integration of the same weights:

Steps	Spatial	Object	Goal	Long
1-step	98.4 (−0.4)	100.0 (+1.0)	97.0 (−1.0)	95.6 (+0.6)
10-step	98.8	99.0	98.0	95.0

Download

Pick the tarball (single file, resumable):

BASE=https://huggingface.co/glowingspirits/let-it-be-simple/resolve/main
aria2c -x16 -s16 --file-allocation=none \
    --dir=$HOME/models -o pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4.tar.gz \
    $BASE/pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4.tar.gz
tar xzf $HOME/models/pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4.tar.gz -C $HOME/models

Then point the repo at them and evaluate:

cd let-it-be-simple
export LBS_CKPT_ROOT=$HOME/models
bash eval/eval_libero.sh pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4 \
     $LBS_CKPT_ROOT/pi0_libero_simvla_tiny_siglip4_head4_w768_50k_h10_a4 \
     libero_10 1        # last argument = integration steps; 1 = one-step decoding

License and third-party terms

The weights in params/ are derived from Google Gemma / PaliGemma pre-trained weights. They are released under the same terms as those models — see LICENSE_GEMMA.txt in the code repository — and the rest of this release is Apache-2.0. The LIBERO benchmark, its datasets and its simulation assets are not included and remain under their own licenses.