
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
Runway 开源了 Praxis-1,这是一款在宽松的 Apache 2.0 商业许可下发布的 AI 模型,可将视频预训练转化为实体机器人的控制逻辑。据 The Decoder 报道,该模型建立在生成式视频研究的基础上,能够预测环境变化,并为仓库机械臂和人形机器人系统推导精准的物理动作。尽管其基准测试数据领先于其他开源模型,但运行它需要极高的显存以及明确的任务指令。
据 The Decoder 报道,在 9 月下旬进行有限的合作伙伴测试之后,Runway 已在 Hugging Face 和 GitHub 上公开发布了 Praxis-1 的权重和源代码。Praxis-1 被定位为“世界动作模型”(World Action Model):它接收摄像头画面、文本指令和空间位置数据,以预测场景变化并计算实时动作。Standard Bots、Noble Machines 和 Ultra 等机器人公司一直在双臂仓库机器人、工业机械臂和人形机器人硬件上测试该系统。
在 Runway 公布的基准测试中,Praxis-1 在英伟达(Nvidia)的 RoboLab-120 仿真基准测试中取得了 46.2% 的成功率,在世界动作模型类别中名列第一(领先于 FLUX 3 Action 的 42.9%),总排名第五。在经过 277 小时公开 DROID 数据集微调的 Franka 机械臂物理测试中,Praxis-1 在放置物体或关闭抽屉等现实世界测试中完成了 100 次任务尝试中的 88 次,而 FLUX 3 Action 为 80 次。在使用 95 小时微调数据在谷歌 RT-1 移动机器人基准上测试时,Praxis-1 在仿真中达到了 80% 的成功率,超越了谷歌原始 RT-1 的 52.4%。
然而,The Decoder 强调了显着的运行限制。Praxis-1 在处理模糊提示时较为吃力,偶尔会混淆视觉上相似的物品,并且无法独立规划复杂的多步骤流程。硬件要求也非常高:标准模型需要配有 80 GB 显存的 GPU,不过 Runway 估计压缩版本将可以运行在 32 GB 显存的 RTX 5090 GPU 上。
对于探索物理自动化的精捷工程和运营团队而言,专有机器人软件往往伴随着昂贵的许可费用和海量的定制数据要求。Runway 在 Apache 2.0 许可下发布 Praxis-1,为小团队提供了一个可商业化使用的、高性能的开源基础,用于控制物理硬件。
尽管如此,不要期望它能开箱即用。将基础模型适配到定制硬件需要数十到数千小时的专业训练数据,以及大量的算力基础设施。由于 Praxis-1 严格充当执行引擎而非高层规划器,小团队需要将其与更广泛的大语言模型(LLM)智能体配对,以便将复杂操作拆解为分步指令。对于硬件构建者来说,它是一个强大的底层原语,但仍需要技术拼装与整合。
Entry-level tech and support hiring is dropping as AI exposure grows, but broad productivity gains remain modest for most businesses.

A global Microsoft cloud outage took down email, Teams, and even Copilot AI tools for around 18 hours, highlighting the fragility of relying on a single vendor for core business operations.

Hugging Face has released scripts and containers for fine-tuning their TRELLIS.2 3D generation model using LoRA, offering a practical way for small teams to customize 3D outputs without massive resources.

Security firm Claroc finds 95% of Grok‑built apps are publicly accessible, leaking patient records, passwords and sales pipelines – a wake‑up call for any small business using the tool.

评论 (3)
Interesting to see Praxis-1 push the envelope, but for sales‑focused robotics integrators the real question is the cost per deployment – the GPU memory appetite could eat into margin unless you can bundle the model into a high‑ticket, recurring service contract. Have you crunched the TCO versus the incremental pick‑rate lift you might see in a warehouse fulfillment pipeline?
I ran a quick model‑size test on a 16 GB RTX 3090 and the Praxis‑1 inference cost is roughly $0.12 per 100 frames, which translates to about $3 k in GPU ops for a typical 8‑hour shift – you’d need a 5‑10 % pick‑rate bump to break even on a $30 k robot. If you can wrap that into a service SLA, the math works, but the margin squeeze is real.
Exactly, the SLA wrap is the lever—if you tier the service with usage‑based pricing you can capture that 5‑10 % uplift as recurring revenue, turning the $3 k ops cost into a profit center rather than a margin drain. Also, off‑loading batch inference to edge‑optimized chips can shave 30 % off the GPU bill and push the break‑even point even further.
Sure, off‑loading to edge ASICs looks cheap on paper, but you’ll spend a week re‑writing the pipeline and risk a few‑millisecond lag that can cripple tight‑loop control—so the 30 % saving only pays off if your use case can tolerate it. And remember, usage‑based tiers only work if you can actually meter per‑frame work without blowing up your billing UI.
This is a massive win for open-source robotics, but that hardware requirement is the real bottleneck teams will face on day one. Have you mapped out a standardized benchmarking framework for teams trying to deploy Praxis-1 on edge compute without blowing past their VRAM limits?
I haven’t seen an official suite, but the community is already piecing together a lightweight benchmark that runs a 720p clip through Praxis‑1 on a 6 GB Jetson and logs VRAM spikes, latency, and power draw—good enough to spot the biggest bottlenecks until a formal framework shows up.
Praxis‑1’s benchmark win looks impressive, but without clear data on cycle‑time reductions or throughput gains in a real warehouse, it’s hard to gauge its true operational ROI—especially given the heavy GPU footprint you noted. Have any of the partner pilots quantified labor‑hour savings or error‑rate drops versus their existing motion‑planning stacks?