MolmoAct2 LIBERO engines for Jetson AGX Thor

Prebuilt TensorRT engines for running allenai/MolmoAct2-LIBERO with vla-edge. This is the configuration used for our 2,000-episode LIBERO evaluation.

With the FP8 vision engine enabled, it scored 97.2% overall, matching the MolmoAct2 paper result within sampling noise. See the performance writeup for the methodology and per-suite results.

Download and verify

This repository is about 10.5 GB.

hf download agents2agents/MolmoAct2-LIBERO-Jetson-Thor \
  --local-dir vla-edge-libero
cd vla-edge-libero
python -c "from vla_edge.backends.tensorrt import artifacts; \
artifacts.check_compatible('.'); artifacts.verify_checksums('.'); \
print('bundle verified')"

Run

vla-edge-serve --embodiment libero --backend tensorrt \
  --engine-dir /path/to/vla-edge-libero/libero --fast-vision

The public API expects two images named image and wrist_image, plus the eight-value LIBERO state vector. The startup warmup executes every compiled stage.

Hardware and contents

These plans require an NVIDIA Jetson AGX Thor Developer Kit with JetPack R39 rev 2.1 and TensorRT 10.16.2.10.

libero/             fixed 704-token engines and FP8 vision option
host/libero/        processor, normalization, embeddings, and flow weights
MANIFEST.json       compatibility requirements and checksums

Serving this bundle is local and does not download the upstream checkpoint.

License

Apache-2.0. See LICENSE and NOTICE.

The plans embed weights from allenai/MolmoAct2-LIBERO, released by the Allen Institute for AI under Apache-2.0. The conversion changes execution, not the checkpoint parameters.