Downloads · 30 days
0
justinchuby/act-aloha-policy-onnx-catalogue
act-aloha-policy-onnx-catalogue is a robotics model from justinchuby. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Real pretrained Action Chunking Transformer policy for AlohaTransferCube-v0. It maps one top-camera observation and 14-DoF robot state to a 100-step, 14-DoF action rollout.
Downloads · 30 days
0
Access
Public
Updated Aug 26, 2026
Repo size
137 MB
Likes
0
Public
Click a slice to open those files.
.onnx137 MB · 100%
From the Hugging Face model README
Real pretrained Action Chunking Transformer policy for AlohaTransferCube-v0. It maps one top-camera observation and 14-DoF robot state to a 100-step, 14-DoF action rollout.
This is the documented fallback for nvidia/Cosmos3-Edge-Policy-DROID. The pinned Cosmos checkpoint does contain a 64-D action tower, but the current Mobius exporter intentionally drops it and installed diffusers lacks the required Cosmos3 action transformer/pipeline. Exact evidence is in COSMOS_BLOCKER.json; no reasoner-only model is presented as a policy.
lerobot/act_aloha_sim_transfer_cube_human@ba73b2766f1371cdc133ca4efb97eb090d7446253c0a209f9fac4d2a57617e686a7f2a2309144ba2hf download justinchuby/act-aloha-policy-onnx-catalogue --repo-type model --local-dir act-aloha-policy
cd act-aloha-policy
pip install 'onnxruntime-gpu==1.28.0' 'numpy==2.3.3' \
'nvidia-cudnn-cu12==9.10.2.21' 'nvidia-cublas-cu12==12.6.4.1' \
'nvidia-cuda-runtime-cu12==12.6.77'
bash run_cuda.sh --request request.npz --output reproduced_output.npz
output.npz contains the real 100×14 action chunk. On H200, steady-state CUDA inference averaged about 12.7 ms; full ONNX-vs-PyTorch maximum absolute difference was below 5e-7. The deterministic zero observation is explicitly a reproducibility fixture, not a simulator episode frame.
inference_metadata.yaml is hashless canonical workflow metadata. Integrity, graph ports, rollout queue semantics, source revisions, license, and performance evidence are stored separately.
Review inference_metadata.annotated.yaml for inline explanations of this package's workflow, tensor/state/cache contracts, and fail-closed omissions. inference_metadata.yaml remains the canonical machine-authored contract; automated validation confirms both files parse to the same metadata object.