AMD publishes a four-demo test of robot training without NVIDIA's CUDA stack

The CUDA moat in Physical AI just got a stress test, and the test came from the chip company that stands to gain if it cracks.

An AMD-aligned team has published a preprint on arXiv walking through four end-to-end robot training and deployment pipelines, all running on AMD's open GPU software stack instead of NVIDIA's proprietary CUDA. CUDA is the proprietary GPU programming toolkit most AI training depends on, and "open" here means AMD's competing ROCm stack plus PyTorch. The setups cover a Franka research arm (a 7-jointed robot arm widely used in academic robotics labs) learning pick-and-place from simulated data, a language-grounded "pick the right one of three objects" task, a real-to-sim pipeline that fuses 3D Gaussian Splatting (a method that turns camera video into a 3D scene mesh) with the Genesis physics engine, and large-scale reinforcement learning for quadruped and humanoid locomotion.

What the paper claims is reproducibility, and that part is concrete. Every demo runs natively on ROCm and PyTorch on RDNA4 (AMD's latest GPU architecture, here the Radeon AI PRO R9700) and RDNA3.5 (the prior generation, here the Radeon PRO W7900), and the authors host the notebooks on a free cloud platform for anyone to rerun. That last part is the unlock: the field has had AMD-claimed Physical AI numbers before, but not a public, no-account-needed place to actually run them.

Physical AI is the industry-coined term for AI that controls robots, drones, and other physical hardware rather than only producing text or images. The paper stakes its claim to "the next major AI frontier" with two industry citations, Jensen Huang at GTC Paris in June 2025 and Dr. Lisa Su at CES 2026, both reproduced as framed inside the preprint rather than verified against independent transcripts, so they read as the authors' framing of an industry mood more than as ground-truth reporting on those executives. The framing, though, matches the public posture of both companies.

The mechanism the paper tests is whether the software stack Physical AI runs on is contestable. Most embodied AI research today runs on NVIDIA's CUDA because the surrounding libraries and tooling, cuDNN, Isaac Sim, Isaac Lab, the robotics simulation stack, were built first and integrated deepest. AMD's bet is that ROCm plus PyTorch plus a small set of robotics libraries can carry the same workload, and the accompanying GitHub repository shows training benchmarks for the R9700 and W7900. If that holds, robotics labs and startups get a second supplier and a price lever on NVIDIA. If not, the lock-in argument wins another data point.

The four demos each answer a different slice of that question. The first uses SmolVLA, a small vision-language-action model that takes camera input and a text instruction and outputs a robot motion, to train a Franka arm on a sim-to-real pick task; the model weights are mirrored on ModelScope. The second adds a language-conditioned "one of three objects" selection problem, testing whether the pipeline can ground written or spoken instructions in a real scene. The third is the most novel of the four: it takes video of a real workspace, reconstructs it as a 3D Gaussian splat, drops that splat into Genesis, and uses the resulting hybrid scene to generate synthetic training data, a way to teach robots from a small amount of real footage. The fourth scales up to a quadruped and humanoid locomotion benchmark, the kind of training run that normally needs large GPU clusters and is where CUDA's tooling is assumed to matter most.

Two limits. The paper is a single-team preprint with no peer review cited, and the AMD-AIM GitHub handle on the accompanying code signals this is ecosystem-native rather than an independent benchmark. The performance numbers the paper alludes to, RDNA4 versus RDNA3.5 versus NVIDIA baselines, sit in supplemental material that the abstract excerpt this story was built from did not include, so the fact-check pass will need to confirm them before any quantitative comparison is published. Both are the reason to treat this as evidence of what is possible, not as a verdict on whether AMD has matched NVIDIA.

What to watch is whether any of the four pipelines gets reproduced outside AMD's cloud by an independent lab, and whether the locomotion results, which is the hardest test, hold up at scale. If they do, the CUDA moat in Physical AI has its first reproducible second supplier. If they do not, this is a well-documented marketing claim rather than a category change.