General Instinct released InstinctFlash, an AGPL-3.0 serving framework that accelerates VLA and world-action models via runtime optimizations and few-step distillation. It speeds a 5B world-action model to real time on Jetson Thor with minimal success-rate loss, and teams at Samsung and Siemens already use it.
A GitHub repo providing optimized inference runtimes for three vision-language-action models on NVIDIA Jetson Thor hardware. An agentic search proposes custom fused kernels and runtime glue, with changes accepted only after passing numerical parity and behavioral tests. It cuts robot inference latency for edge robotics developers without distillation or pruning.
A GitHub repo providing optimized inference runtimes for three vision-language-action models on NVIDIA Jetson Thor hardware. An agentic search proposes custom fused kernels and runtime glue, with changes accepted only after passing numerical parity and behavioral tests. It cuts robot inference latency for edge robotics developers without distillation or pruning.
A new arXiv paper introduces FoldQuantVLA, a post-training quantization framework for vision-language-action models that runs four-bit weights and activations natively via custom TensorRT plugins. Tested on GR00T and π0.5 checkpoints, it delivers 1.2–1.5x speedups while a mixed W4A4/W8A8 configuration restores robot task success rates.
A new arXiv paper introduces FoldQuantVLA, a post-training quantization framework for vision-language-action models that runs four-bit weights and activations natively via custom TensorRT plugins. Tested on GR00T and π0.5 checkpoints, it delivers 1.2–1.5x speedups while a mixed W4A4/W8A8 configuration restores robot task success rates.
Researchers propose Jetson-PI, a method to deploy Vision-Language-Action (VLA) models on low-power Jetson Orin devices for real-time robot control. It addresses inference latency and control frequency issues by using asynchronous inference with foresight alignment to mitigate perception-execution misalignment and long reaction times.
Researchers propose Jetson-PI, a method to deploy Vision-Language-Action (VLA) models on low-power Jetson Orin devices for real-time robot control. It addresses inference latency and control frequency issues by using asynchronous inference with foresight alignment to mitigate perception-execution misalignment and long reaction times.
Researchers present rMuscle, a real-time Vision-Language-Action inference framework that exploits similarity across repeated robot executions via a dual-phase muscle-memory cache. It delivers 1.29-1.42X speedups on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and physical tasks while preserving original success rates, addressing latency bottlenecks that affect robot responsiveness in factory settings.
Researchers present rMuscle, a real-time Vision-Language-Action inference framework that exploits similarity across repeated robot executions via a dual-phase muscle-memory cache. It delivers 1.29-1.42X speedups on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and physical tasks while preserving original success rates, addressing latency bottlenecks that affect robot responsiveness in factory settings.
Researchers propose VQVLA, an algorithm-hardware co-design framework that accelerates Vision-Language-Action (VLA) model inference by exploiting weight similarity and execution dynamics. The key innovation is MotionVQ, a motion-aware vector quantization scheme that dynamically reduces memory and computation redundancy. This work targets the high inference latency that limits real-time deployment of VLA models on GPUs.
Researchers propose VQVLA, an algorithm-hardware co-design framework that accelerates Vision-Language-Action (VLA) model inference by exploiting weight similarity and execution dynamics. The key innovation is MotionVQ, a motion-aware vector quantization scheme that dynamically reduces memory and computation redundancy. This work targets the high inference latency that limits real-time deployment of VLA models on GPUs.