MolmoAct2-Jetson-Thor

MolmoAct2 BimanualYAM engines for Jetson AGX Thor

Prebuilt TensorRT engines for running the original allenai/MolmoAct2-BimanualYAM checkpoint with vla-edge. The fastest configuration produces one 30-action chunk in 113.4 ms on our reference Thor, down from 611.3 ms with the PyTorch backend.

See the performance writeup for the methodology and full results.

<p align="center"> <a href="https://agents2agents.ai/blog-assets/molmoact2-thor-cube-basket.mp4"> <img src="https://agents2agents.ai/blog-assets/molmoact2-thor-cube-basket-poster.jpg" width="720" alt="MolmoAct2 running on Jetson AGX Thor and placing a Rubik's cube in a black basket with two YAM arms"> </a> <br> <a href="https://agents2agents.ai/blog-assets/molmoact2-thor-cube-basket.mp4">Watch the real-time robot demo</a> </p>

Download and verify

This repository is about 11.6 GB. It contains only the fastest BimanualYAM deployment, not the separate LIBERO or dynamic-prompt engines.

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

Checksum verification reads the full bundle and can take a few minutes.

Run

Install vla-edge and complete its Jetson Thor prerequisites, then start the server:

vla-edge-serve --embodiment bimanual-yam --backend tensorrt \
  --engine-dir /path/to/vla-edge-thor/yam-champion --fast-vision

The startup warmup executes every compiled stage. From another terminal:

curl -f http://127.0.0.1:8202/act

The response should report "backend":"tensorrt", "rtc":true, and the bimanual-yam embodiment. The full robot setup is in the Bimanual YAM example.

Hardware and contents

These plans were built for:

• NVIDIA Jetson AGX Thor Developer Kit, sm_110a, aarch64
• JetPack R39 rev 2.1
• TensorRT 10.16.2.10

MANIFEST.json records the hardware requirements, file sizes, and SHA-256 checksums. vla-edge checks the machine before loading model state.

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

The default LLM prefill is bf16. Vision is fp16 by default and FP8 with --fast-vision. The compiled action expert is fp16. A portable TensorRT action engine is included as a fallback when the installed PyTorch ABI does not match the compiled flow package.

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

Other engine sets

• Dynamic-prompt BimanualYAM
• LIBERO

License

Apache-2.0. See LICENSE and NOTICE.

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