A physical AI model can score well in a simulator and still fail on a real robot, vehicle or test rig. Richard Ahlfeld argues that this happens because a simulation contains only the physics, sensors and scenarios its builders put in. Ahlfeld is CoreWeave’s SVP for Physical and Scientific AI. He made the case in the May 14, 2026 episode of CoreWeave’s AI Cloud Essentials podcast, “Getting Physical with AI,” hosted by Ritu Jyoti. His conclusion is not to drop simulation. He says to use it for scale and to use physical evidence to check whether its assumptions hold.
The core claim: simulation is necessary but bounded
Physical AI has to connect perception to action in real environments, where sensor input, timing, hardware and safety limits all matter. Simulation and synthetic data help by generating many scenarios quickly and exposing edge cases. They are only as credible as the simulated physics behind them.
As an Amazon Associate I earn from qualifying purchases.
Ahlfeld put it bluntly in the interview: “simulations are good, but they will never be as good as the real world.” That is a conversational statement, not a formal scientific finding. Read it as a warning about where evidence comes from, not as a rejection of simulation.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWho is speaking and what the interview covered
CoreWeave’s episode overview frames the conversation around simulation, testing, engineering decision-making, and automotive, manufacturing and robotics applications. Ahlfeld describes a path from aerospace engineering and physics-informed AI research into industrial engineering work. His examples include aircraft engines, NASA work and his earlier company, Monolith.
#1 Best Overall
- AI-Powered Raspberry Pi Robot Dog — PiDog: Powered by Raspberry Pi (5/4B/3B+/3B/Zero 2W), OpenClaw, and multi-LLMs like ChatGPT, Gemini, Grok, DeepSeek, Qwen & Ollama. With 12 servos, camera, gyroscope, hearing & touch sensors, PiDog can see, listen, talk, move, and interact intelligently. Supports OpenCV, MediaPipe, TTS & STT, app control, FPV & Python. A great STEM robotics gift for students, makers & tech enthusiasts—perfect for birthdays and holidays. (Raspberry Pi not included)
- Realistic Dog-like Movements: PiDog's 12 powerful servos enable 32 dog-like actions, including walking, sitting, standing, shaking its head, wagging its tail, and performing playful tricks, closely mimicking a real dog and providing an engaging experience. This is an AI development robot product designed for engineers, suitable for ages 15 and above
- Rich Sensor Suite for Interactive Experiences: PiDog features ultrasonic, touch, gyroscope, sound, camera, speaker and microphone. These provide it with advanced hearing, vision, and touch, enabling it to see, detect obstacles, respond to touch, and recognize sounds, making interactions highly engaging
- AI-Powered Interactions with OpenClaw & Multi-LLMs. PiDog combines voice, vision, and gesture recognition for immersive AI experiences. Powered by OpenClaw and multi-LLMs like ChatGPT, Gemini, Grok, DeepSeek, Qwen, Doubao, and Ollama (local LLMs), it can understand questions, respond naturally through TTS & STT, recognize math problems, interpret hand gestures, and hold smart conversations. OpenClaw also enables customizable AI behaviors and personalized robotics development, helping users create their own intelligent robotic companion
- Comprehensive Learning Resources and Support: PiDog offers detailed online documentation, video tutorials, prompt technical support, and an active forum community, ensuring beginners can easily complete all projects and enjoy a great experience
The published transcript contains automatic-transcription errors in several names and phrases. This article relies on the official overview and a later interview for role and company context. It does not reproduce garbled transcript wording.
What is the sim-to-real gap? The water bottle example
Ahlfeld’s main illustration is simple. In a simulation, a plastic bottle may behave as a rigid object. In a lab, a real bottle can deform or crumple when a gripper handles it. A robot trained only on the rigid version has learned something that is wrong for the task, however fast the training ran.
His remedy is to put a robot in a lab to practise with a real water bottle and collect physical feedback. The lesson is that a fast simulation can be accurate on many properties and still omit the one that decides the outcome.
Free tools Windows power users keep installed
One-click scans. No signup required.
What simulation and synthetic data do well
- Coverage: teams can vary lighting, layouts, object positions and traffic conditions far beyond what they could stage physically.
- Labels: generated data can come with ground truth attached.
- Safety and cost: dangerous or expensive situations can be explored without risking hardware or people.
- Iteration speed: many configurations can be compared quickly.
CoreWeave’s physical AI materials present simulation-generated training data as one stage in a larger workflow. The other stages are multimodal sensor fusion, inference, retraining and staged validation.
Rank #2
- Optimized AI Arm Kit for LeRobot & Hugging Face Projects – The SO-ARM101 is an upgraded low-cost robotic arm servo motor kit designed for AI robotics enthusiasts and developers. Fully compatible with LeRobot and Hugging Face frameworks, it supports imitation learning and reinforcement learning, making it ideal for real-world robotics applications. (3D-printed parts not included.)
- Enhanced Wiring & Performance – Compared to the SO-ARM100, the SO-ARM101 features improved wiring to prevent disconnection at joint 3 and eliminates range-of-motion limitations. The leader arm uses optimized gear ratio motors for smoother performance—no external gearboxes required.
- Real-Time Leader-Follower Functionality – New real-time tracking allows the leader arm to follow the follower arm, enabling human intervention and correction during reinforcement learning (RL) training. Perfect for hands-on AI robotics development and research.
- Open-Source, DIY-Friendly & Nvidia-Compatible – Developed by TheRobotStudio, this open-source AI Arm kit integrates seamlessly with the LeRobot platform, offering PyTorch-based datasets, simulation, training, and deployment tools. Fully compatible with Nvidia Jetson edge devices, including reComputer Mini J4012 Orin NX 16 GB.
- Comprehensive Learning Resources – Includes detailed open-source assembly and calibration guides, testing tutorials, and deployment instructions. From wiring to AI training, get everything you need to start building, teaching, and optimizing your robotic arm for grasping and placing tasks.
What physical tests catch that simulation tends to miss
Physical testing gives evidence about the actual system. In a September 2026 interview with IZON, Ahlfeld pointed to several areas that are hard to capture completely in simulation:
- robot grasping mechanics, including contact and material deformation;
- liquids;
- chaotic human behavior;
- sensor or hardware failure.
These are his examples. They do not mean simulation can never model such things. They mean each needs a credible model, and credibility has to be checked against real observations.
A validation loop instead of a winner
The workflow that follows from his argument is circular, not a choice between synthetic and real data:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Observe real systems and simulated environments.
- Curate real data and generate synthetic data.
- Train models.
- Evaluate behavior, including against held-out physical results.
- Deploy in stages.
- Feed new outcomes back into the next cycle, correcting the model and the simulation.
Simulation supplies scalable coverage. Physical evidence tests whether the model’s assumptions hold on the target hardware and sensors.
Rank #3
- Raspberry Pi AI Robot: powered by Raspberry Pi (5/4B/3B+/3B/Zero 2W), features 12 servos and sensors for vision, hearing, and touch. Integrated with ChatGPT-4o, it responds to complex queries. With app control and FPV, users can manage and see its view in real-time. It supports Python programming
- Realistic Movements: 12 powerful servos enable 32 actions, including walking, sitting, standing, shaking its head, wagging its tail, and performing playful tricks, closely mimicking a real and providing an engaging experience
- Rich Sensor Suite for Interactive Experiences: features ultrasonic, touch, gyroscope, sound, camera, speaker and microphone. These provide it with advanced hearing, vision, and touch, enabling it to see, detect obstacles, respond to touch, and recognize sounds, making interactions highly engaging
- Engaging Interactions with ChatGPT-4o: with ChatGPT-4o enables voice interactions and visual recognition, making it smarter and more responsive. Users can have natural conversations, solve math problems via the camera, and interpret gestures, creating diverse and fun interactions
- Comprehensive Learning Resources and Support: offers detailed online documentation, video tutorials, prompt technical support, and an active forum community, ensuring beginners can easily complete all projects and enjoy a great experience
The Nissan example, and its caveat
Ahlfeld described using historical hardware and physical test data to predict what would happen in real chassis tests. He said this let Nissan reduce testing across its chassis by 17%. He immediately qualified it: many tests were safety-critical and could not be omitted.
Treat this as his reported case-study result. We found no independent audit or academic or regulatory study of the figure, and it should not be read as a universal effect. It also supports the same theme. The prediction was built on real test data, so physical evidence reduced future physical testing instead of replacing the need for it.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What a simulation count does and does not prove
CoreWeave’s blog gives several workload examples. They are company-reported throughput figures for specific setups, not cross-platform benchmarks or deployment guarantees.
| Workload (reported by CoreWeave, 2026) | Tool | Scale and time |
|---|---|---|
| Robotic manipulation | MuJoCo | 4,800 simulations in 85 minutes |
| Randomized warehouse scenes | NVIDIA Isaac Sim | 10,000 samples in 21 minutes |
| Larger Isaac Sim run | NVIDIA Isaac Sim | 113,000 simulations in just under eight hours |
| Autonomous vehicles | CARLA | 1.25 million simulations in about 12 hours |
| Driving rollouts | AlpaSim | More than 1,600 rollouts in under four hours |
These numbers show how fast scenarios can be produced. They do not show that the scenarios were realistic, that the physics matched the target system, or that the model transfers safely. CoreWeave itself says the AlpaSim result is for failure triage before road testing and is not safety certification on its own.
Rank #4
- 【End-to-End Imitation Learning】Hiwonder SO-ARM101 robot arm is an embodied intelligent hardware platform compatible with the Lerobot open-source framework. It provides developers with streamlined access to shared code, templates, and pre-trained models to explore the latest advancements in AI research.
- 【Dual-Camera Vision System】Equipped with both a gripper-mounted camera and an external camera, the system supports both precise manipulation and environmental awareness for accurate imitation learning.
- 【Hiwonder High-Performance Bus Servos】Featuring 12 high-torque bus servo motors with magnetic feedback, the Hiwonder SO-Arm101 robotic arm delivers smooth, stable motion, eliminating issues like power deficiency and jitter.
- 【Professional Control & Debugging】Integrated with the Hiwonder BusLinker V3.0 debugging board, the system supports servo scanning, real-time status monitoring, and trajectory control. The professional PC software simplifies device calibration and debugging, making it accessible for both researchers and hobbyists.
- 【Open-Source Compatibility】The SO-ARM101 robotic arm is designed to be fully compatible with the LeRobot open-source project. We acknowledge the contributions of the open-source community; all trademarks and copyrights belong to their respective owners.
How to decide between synthetic data and physical tests
The sources do not rank vendors or products, so use these axes to judge any approach:
| Question | What to look for |
|---|---|
| Scenario coverage and throughput | Does simulation reach rare conditions you could not stage? |
| Fidelity | Does the simulator model the behavior that decides success, such as deformation, contact or liquids? |
| Real-world data | Is there enough, and is it good enough, to calibrate and check the model? |
| Target hardware and sensors | Was performance measured on the actual system, not an idealized one? |
| Rare failures | Can you detect them and feed them back into training? |
| Operational latency | Does the model meet real-time constraints when deployed? |
| Deployment evidence | What physical and staged results are required before release? |
In practice, lean on simulation where the physics is well understood and the goal is breadth. Insist on physical tests where materials, contact, liquids, people or hardware faults drive the outcome, and wherever safety is at stake.
Company context and how to read the claims
CoreWeave’s September 2026 description of its Physical AI Field Engineering service says engagements begin with an on-site scoping workshop. They can then include simulation infrastructure, analysis of test and sensor data, and building applications or models for customer workflows. That is the company’s own description of its approach, not independent validation. A September 10, 2026 IZON report describes CoreWeave’s physical AI customers across robotics, autonomous vehicles and industrial applications. Its figures come from that report and company representatives.
In the Field Engineering announcement, Ahlfeld, there titled Senior Vice President of Physical AI, said: “Engineering teams don’t adopt a new method because a vendor proved it once in a demo. They adopt it once they’ve seen it hold up on their own systems.” That is an executive’s account of customer adoption, and it points the same way as his simulation argument. Evidence from your own system is what counts.
Remember also that CoreWeave sells compute and engineering services for this kind of work. That gives it a commercial interest in simulation at scale, which is part of why the caveats it states itself are worth noting.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
