NVIDIA Isaac、CoreWeave Physical AI Launchables 與 AWS 及 NVIDIA 工具鏈方案比較

Within seven days, three different companies handed developers a physical AI toolchain. NVIDIA expanded Isaac at GTC Berlin on October 21. CoreWeave shipped one-command robotics templates the same day. AWS and NVIDIA open-sourced a reference Physical AI toolchain a week earlier, on October 14.

If you build software and not robots, you're wondering why you should care. Fair question. Here's the honest version: nobody is asking you to buy a humanoid. These launches matter because they're all selling the same thing — the boring, expensive middle layer of robot development: data generation, simulation, and evaluation. And they're giving it away so you'll spend money on their compute.

Limits first, because that's the deal here: I haven't spun up a Jetson Thor on a real humanoid. Nobody outside a handful of labs has. This is a docs-and-announcements-verified breakdown — what each launch actually ships, what the catches are, and what I'd do this week if I were you.

Quick verdict

Three plays, three strategies:

• NVIDIA Isaac expansion — the deepest stack. A new Isaac OS runtime, Isaac Lab 3.0 for robot learning, and open reference workflows pairing ChatGPT Astra + Isaac + Cosmos on a DGX Station. If you're already in NVIDIA's orbit, this is the fullest thing on the table.
• CoreWeave Physical AI Launchables — the laziest start. One-command templates for CARLA, MuJoCo, NVIDIA AlpaSim and two more tools, spanning data generation, grading, and policy training. Built for teams that already rent GPU time and hate setup.
• AWS + NVIDIA toolchain — the enterprise lane. Open-source reference architectures wiring Isaac Sim, Isaac Lab, GR00T, and Cosmos into SageMaker, EC2, and IoT Greengrass. If your company lives on AWS, this is the path of least paperwork.

The pattern across all three: free software, metered compute. Keep that sentence in your head for the rest of the article. It explains everything.

Side by side

Play 1: NVIDIA's Isaac expansion — the deepest stack, and the deepest moat

Jensen Huang's line from Berlin: "General-purpose robotics is advancing at extraordinary speed — the ChatGPT moment for robotics is fast approaching."

That's a quote, not a fact. File it accordingly. The actual shipping pieces are worth your attention, though:

Isaac OS runtime — a runtime layer for robots. Read it as NVIDIA trying to own the "operating system" position in robotics the way CUDA owns GPU compute. If that framing lands, every robot team eventually pays the NVIDIA tax, just like every ML team does today.

Isaac Lab 3.0 — the robot learning framework. This is the piece most developers can actually touch: it's where you train and evaluate robot policies in simulation before anything touches physical hardware.

Reference workflows: ChatGPT Astra + Isaac + Cosmos on DGX Station — data generation, training, and evaluation in one documented loop. Synthetic data from Cosmos, robot frameworks from Isaac, orchestrated with Astra, running on a DGX Station. It's the first time the whole loop has been packaged as a single documented path instead of five forum threads duct-taped together.

Europe's robotics names — ABB, KUKA, NEURA Robotics, Universal Robots — are building on it, per the announcement. That's the social-proof slide.

The catch. Three of them. First, "open robotics platform" is doing a lot of work — the openness lives inside NVIDIA's ecosystem. You can build freely, but every road leads back to their silicon and their frameworks. Second, the reference hardware story (more on the Agile ONE SN below) is still announcement-grade: no price, no third-party teardown, no independent setup-time verification. Third, the thing nobody says out loud: Isaac Lab 3.0 is free. The GPU hours to train a policy in it are not.

Play 2: Agile ONE SN — the reference humanoid (and why you shouldn't buy one yet)

Same event, separate announcement: Agile Robots' Agile ONE SN, a full-stack reference humanoid preconfigured with NVIDIA Isaac GR00T and Jetson Thor edge AI. The spec sheet: 1.40m tall, 40kg, 31 degrees of freedom, 3kg payload per arm, dual WUJI Hand 2 five-finger hands (20 DoF each, 1kHz force control, tactile sensing). The compute: Jetson Thor with 2,070 FP4 teraflops and 128GB unified memory. The claim: setup time from weeks to days.

"Reference" is the key word. This isn't a product review; it's a template for what a GR00T-powered humanoid should look like. For developers, the value is that there's now a documented hardware target to build against — write your stack for the reference design, and your code ports to anything that follows the template.

The catch. The announcement gives no price, and a 40kg humanoid with a Jetson Thor brain and 1kHz force-controlled hands is not going to be a weekend purchase. "Weeks to days" is a vendor claim with zero independent verification so far. And 31 degrees of freedom means 31 things that need calibration, maintenance, and debugging — the reference design answers the architecture question, not the reliability one. I'd wait for a third party to actually stand one up and report back before budgeting anything.

Play 3: CoreWeave Physical AI Launchables — the laziest start

CoreWeave's angle is different: no platform, no hardware, just templates. Launchables are tested, one-command workload templates for robotics, AV, and simulation teams. Five tools at launch, including CARLA, MuJoCo, and NVIDIA AlpaSim. They cover the physical AI loop end to end: data generation, evaluation/grading, policy training. Every tool ships a blank template you point at your own scenes and data, then run on reserved or burst capacity.

The genuinely interesting part isn't the one-command bit. It's the agentic skills: generative AI data augmentation (video, image, text), auto-labeling, data curation and retrieval. Data is the actual bottleneck in robotics — real-world robot data is expensive, slow, and miserable to label. Anything that attacks the data problem is attacking the real problem.

The catch. "One command" gets you the template, not the pipeline. Your scenes, your data, your labels, your evaluation criteria — all still yours to build. And CoreWeave is a GPU cloud: the templates are the free sample; the meter runs on reserved or burst GPU capacity. That's a fair trade if you're already renting GPUs, but don't mistake a template catalog for a managed service.

Play 4: AWS + NVIDIA toolchain — the enterprise lane

A week earlier (Oct 14), AWS and NVIDIA open-sourced a Physical AI toolchain: Isaac Sim, Isaac Lab, GR00T, and Cosmos wired into SageMaker, EC2, and IoT Greengrass, with reference architectures and deployment automation. Open-source, deployable now.

This is the one your company's platform team will actually pick. Not because it's the best technology — it's a reference architecture, same family as NVIDIA's — but because it arrives inside the AWS procurement boundary. Greengrass for edge deployment is the tell: this is built for factories and warehouses that already run AWS, not for garage robotics.

The catch. Open-source reference, AWS-priced runtime. The architectures are free; the SageMaker training jobs and the EC2 simulation fleet are billed like everything else on AWS. And "reference architecture" means you're still doing integration work — it's a starting point, not a product.

What maps to what

Cut through the announcements and here's what's actually on the table:

• Simulation: Isaac Sim (NVIDIA/AWS lane) vs MuJoCo vs CARLA (AV lane) vs AlpaSim. The field hasn't settled. MuJoCo is the research default; Isaac Sim is the commercial push; CARLA owns the driving corner.
• Learning: Isaac Lab 3.0 is the framework to learn if you want one bet. It's where NVIDIA, AWS, and the reference humanoid all converge.
• Data: Cosmos (synthetic, NVIDIA) vs CoreWeave's agentic augmentation skills vs your own real-world collection. Synthetic data is the bet every vendor is making, because real robot data doesn't scale.
• Compute: DGX Station (own the box) vs CoreWeave burst (rent by the hour) vs AWS (you're already there). All three monetize the same way: free tools, metered GPUs.
• Edge: Jetson Thor is the new reference brain. 2,070 FP4 teraflops in an edge module is the number that makes onboard robot inference plausible without a server rack.

What I'd actually do this week

Five things, in order of effort:

1. Install Isaac Lab 3.0 and run a sim policy. Costs nothing, teaches you the loop every one of these launches is built around. If you learn one thing this week, learn this.
2. If you rent GPUs already: read CoreWeave's Launchables templates before writing your own infrastructure. The blank templates plus the auto-labeling and data-curation skills are worth stealing as patterns even if you deploy elsewhere.
3. If you're AWS-native: skim the Oct 14 reference architectures. The SageMaker + Greengrass wiring is the part your platform team will ask about. Now you have an answer.
4. Don't buy the humanoid. Agile ONE SN has no price and no independent verification. Admire the spec sheet (1kHz force control is genuinely impressive on paper), budget nothing.
5. Price the compute before you train. Every one of these plays gives away the software and bills the GPUs. Run the numbers on a training run before you commit to a stack — the template is free, the policy is not.

The part nobody's saying

All three launches, plus the reference humanoid, add up to one strategy: make the software free and the tooling excellent, then collect on compute. NVIDIA sells the GPUs. CoreWeave rents them. AWS bills them. The developer experience genuinely improved this week — a documented data-to-policy loop is real progress — but improved DX is also the funnel.

That's not a reason to avoid any of it. CUDA won the same way, and it was good for developers for a long time. Just know which game you're playing: you're getting world-class free tooling in exchange for eventually paying someone for GPU hours. Budget accordingly, keep your data portable, and don't build anything that only runs on one vendor's cloud.

Are any of you actually running sim-to-real anywhere yet, or is your physical AI exposure still YouTube videos? And if you've touched Isaac Lab, MuJoCo, or CARLA this year: which one would you bet a side project on, and why? Genuinely curious where the practitioner consensus is landing.

Sources: NVIDIA GTC Berlin Isaac expansion (GlobeNewswire, Oct 21, 2026); Agile Robots Agile ONE SN (GTC Berlin, Oct 21, 2026); CoreWeave Physical AI Launchables (Business Wire, Oct 21, 2026); AWS + NVIDIA open-source Physical AI Toolchain (Oct 14, 2026). Research-based analysis — no hands-on hardware testing claimed.