General Instinct 推出了 InstinctFlash，这是一个专为机器人模型设计的高性能推理框架，采用 AGPL-3.0 开源许可证发布1。该框架在 NVIDIA Jetson Thor 硬件上实现了 1.2 倍至 7.9 倍的运行时加速，结合扩散调度器优化技术可进一步达到 33.78 倍的加速效果1。
在 Robotwin2.0 任务集的 1,153 个评估案例中，LingBot-VA 模型在 InstinctFlash 上优化至 2 个视觉步骤和 4 个动作步骤后，达到了 90.5% 的成功率1。该框架支持包括 pi0.5 和 NVIDIA Cosmos Policy 在内的 8 个 VLA/world-action 模型系列，可运行在 RTX 4090/5090 和 Jetson Thor 等硬件平台上1。目前三星、西门子等公司已将其用于生产环境1。
General Instinct has released InstinctFlash, an open-source inference framework designed for robot models and released under the AGPL-3.0 license.1 The framework achieves significant performance improvements on Jetson Thor hardware, delivering runtime acceleration ranging from 1.2x to 7.9x, with combined optimizations using a diffusion scheduler reaching up to 33.78x acceleration.1
In evaluation on the Robotwin2.0 task set across 1,153 episodes, the LingBot-VA model achieved a 90.5% success rate with 2 visual and 4 action steps, compared to a baseline success rate of 92.1% using 25 visual and 50 action steps.1 InstinctFlash supports eight VLA and world-action model families, including pi0.5 and NVIDIA Cosmos Policy, and is compatible with RTX 4090/5090 GPUs and Jetson Thor hardware.1 The framework has already been adopted in production environments by Samsung, Siemens, and other robotics startups.1
评论
还没有评论，欢迎留下第一条。