- Category
- Developer Tools
- Rank
- No. 1748Tools index
Previous survey · No. 1694 ·
- Type
- TOOL
- Use case
- Models: Train & Run
- GitHub
- 12 stars
- Date
About
vla-edge provides optimized inference runtimes for three vision-language-action models (ABC-VLA, MolmoAct2, and pi-0.5) running on NVIDIA Jetson Thor hardware. Using an agentic search process to propose custom fused CUDA kernels and runtime glue, alongside CUDA graphs, TensorRT, and a lossless weight decoder, it cuts MolmoAct2 inference from 611ms to 113ms and ABC-VLA inference to 32.4ms versus 63.4ms for a plain TensorRT conversion, without distillation, pruning, or sub-fp16 quantization.
At a glance
vla-edge runs supported vision-language-action policies on the Jetson AGX Thor. It is worth investigating if your robot already uses one of its supported checkpoints and the time to produce the next chunk of actions is limiting your control loop.
Good fit
Robotics teams evaluating MolmoAct2, ABC-VLA or π0.5 on the Jetson AGX Thor.
Consider something else when
Look elsewhere if you need a drop-in accelerator for an arbitrary policy or a safety guarantee for a physical robot.
Start here
Pick the recipe that matches your checkpoint. Use a recorded input to compare action outputs and per-chunk latency with your existing backend before connecting actuators.
Cost and requirements
The documented Thor setup requires compatible NVIDIA hardware and a matching PyTorch or TensorRT toolchain. Weights and compiled engines come separately. Review the checkpoint loader before allowing remote model code.
Limits to know
The published latency figures are vendor measurements for specific models and precisions. They measure time per action chunk, which does not establish control-loop latency or safe motion on your robot.
Sources checked 2026-11-05. Source-based assessment.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.