Downloads · 30 days
0
lair-nyu/yor_icl_victr_vision_interp_canonical_ctx3cam-step5000
yor_icl_victr_vision_interp_canonical_ctx3cam-step5000 is a machine learning model from lair-nyu. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
RICL-pi0.5: RICL's action-interpolation mechanism (blending the model's flow-matching prediction toward the top-1 vision-retrieved chunk's action, weighted by exp(-lamda vision distance), lamda=3.0) ported onto the pi…
Downloads · 30 days
0
Access
Public
Updated Sep 14, 2026
Repo size
12.4 GB
Likes
0
Public
Click a slice to open those files.
Other12.4 GB · 100%
From the Hugging Face model README
RICL-pi0.5: RICL's action-interpolation mechanism (blending the model's flow-matching prediction toward the top-1 vision-retrieved chunk's action, weighted by exp(-lamda * vision distance), lamda=3.0) ported onto the pi0.5 backbone (3-camera k=1 canonical context encoding). Used as the RICL-pi0.5 baseline in the paper's comparison against the three VICTR arms. Checkpoint from a still-in-progress training run.
Only params/ (model weights), assets/ (norm_stats.json) and
_CHECKPOINT_METADATA are included -- train_state/ (optimizer state, ~1.5x
the size of params, only needed to resume training) is dropped, matching the
existing checkpoint-repo convention in this project.
Loading this checkpoint requires the training config it was built from
(yor_icl_victr_vision_interp_canonical_ctx3cam in openpi/src/openpi/training/config.py) -- this repo only
carries weights + norm stats, not the model architecture or data transforms.
With that repo checked out:
from openpi.policies import policy_config
from openpi.training import config as _config
policy = policy_config.create_trained_policy(
_config.get_config("yor_icl_victr_vision_interp_canonical_ctx3cam"),
"<local snapshot_download of this repo>",
)