## Does Humanoid Robot Training Actually Require NVIDIA CUDA?

A new paper from Qing Yang, Xun Wang, Ziguan Wang, Zhenjiang Li, Hongqiang Wang, and Dongdong Weng says no — and demonstrates an end-to-end [Physical AI](https://humanoidintel.ai/glossary/physical-ai) stack running entirely on AMD hardware. Published today on arXiv (2607.22997), the work presents a fully AMD-accelerated pipeline spanning data-center training silicon, Radeon PRO simulation and rendering GPUs, and Ryzen AI edge compute, all unified under the open ROCm software stack. The paper's headline claim is direct: training and deploying [Vision-Language-Action Model](https://humanoidintel.ai/glossary/vision-language-action-model) (VLA)-based manipulation policies does not require a CUDA-locked ecosystem.

The pipeline covers four progressive demonstrations: a [Sim-to-Real Transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) manipulation pipeline using SmolVLA deployed on a physical Franka arm; a semantic language-grounded object-selection task; a Real2Sim synthetic-data generation pipeline fusing 3D Gaussian Splatting reconstructions with the Genesis physics engine; and large-scale reinforcement learning for both quadruped and **humanoid locomotion** benchmarked across multiple hardware platforms. All pipelines run natively on ROCm and PyTorch on RDNA4 (Radeon AI PRO R9700) and RDNA3.5 (Radeon PRO W7900) hardware, with reproducibility available on the free Radeon Cloud Platform.

For the humanoid robotics industry, where sim-to-real transfer and VLA policy training represent core infrastructure costs, a credible CUDA alternative has significant supply-chain and cost implications.

---

## What the Pipeline Actually Does

The paper's architecture is structured as a progression from simulation to real-world deployment and back — the "Real2Sim2Real" loop that gives the paper its name.

**Demo 1: SmolVLA on Franka.** The baseline demonstration trains a VLA policy using SmolVLA — a compact, open-weight VLA model — and deploys it on a physical Franka robotic arm. The source material does not specify success rates or task complexity beyond the hardware setup, so performance comparisons to CUDA-trained equivalents remain an open question.

**Demo 2: Language-Grounded Object Selection.** A "one-of-three" semantic task requires the system to parse a natural-language instruction and select the correct object from a set of three. This is a meaningful test of [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) for VLA policies — the system must ground language to perception without task-specific fine-tuning on every possible object combination.

**Demo 3: Real2Sim with 3D Gaussian Splatting.** This is arguably the most technically novel contribution for the simulation pipeline community. The authors fuse 3D Gaussian Splatting (3DGS) reconstructions of real-world scenes with the Genesis physics engine to generate synthetic training data. The Real2Sim direction — scanning real environments to build high-fidelity simulation assets — directly addresses the data bottleneck that hampers [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) policy training. Rather than manually authoring simulation environments, the pipeline reconstructs them from RGB or RGBD captures of the actual deployment space.

**Demo 4: Humanoid and Quadruped Locomotion RL.** Large-scale reinforcement learning for humanoid locomotion is benchmarked across multiple hardware platforms. The source does not provide specific benchmark scores, training times, or comparative CUDA vs. ROCm wall-clock figures — a notable gap for a paper making hardware-agnostic claims. That said, the cross-platform benchmarking framing suggests the authors are positioning this as an infrastructure paper rather than a locomotion performance paper.

---

## Why the AMD Angle Matters Strategically

The humanoid robotics stack currently has deep NVIDIA dependencies at every layer: Isaac Sim for simulation, CUDA for training compute, Jetson for edge inference in some platforms. This isn't accidental — NVIDIA invested heavily in developer tooling, and CUDA's decade-plus head start in ML frameworks created genuine lock-in through library compatibility rather than explicit exclusivity.

Jensen Huang's framing at GTC Paris in June 2025 ("the next big thing is Physical AI, AI with a body") and AMD CEO Lisa Su's CES 2026 remarks ("we're entering the world of Physical AI ... this is where AI enters the real world"), both cited in this paper, signal that Physical AI is now a platform war, not just an application domain. AMD is positioning ROCm as the open alternative to CUDA — and this paper is an existence proof that the stack is functional for humanoid-relevant workloads.

**The skeptical read:** The paper does not provide head-to-head CUDA vs. ROCm performance benchmarks with specific numbers. "Reproducible" and "performant" are different claims. For humanoid robotics teams evaluating whether to move sim-to-real training infrastructure off NVIDIA hardware, the absence of quantitative throughput comparisons is a real limitation. A team running 10,000 parallel RL environments for whole-body control needs to know GPU-hour costs, not just that the pipeline runs.

**The constructive read:** Infrastructure papers often matter more for ecosystem development than their raw numbers suggest. If AMD's ROCm stack can run the full VLA training and deployment loop — data generation, policy training, edge inference — without CUDA, it creates genuine competition in the compute supply chain. For humanoid robotics startups burning through GPU budgets on NVIDIA hardware, an alternative matters even if the performance parity is 80-90% rather than 100%.

---

## Humanoid Locomotion RL: The Scope Question

This paper covers quadruped locomotion alongside humanoid locomotion in Demo 4. For scope purposes: the humanoid locomotion benchmarking component is directly relevant to [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) research for bipedal platforms. Large-scale RL for humanoid locomotion is one of the most compute-intensive workloads in the field — training bipedal gaits requires simulating contact dynamics, balance recovery, and terrain adaptation at scale. Any viable alternative to CUDA-based clusters for this workload class is industry-relevant.

The Genesis physics engine integration is worth watching. Genesis has emerged as a fast, GPU-accelerated physics simulator that has attracted attention from several humanoid-adjacent research groups for its parallel simulation throughput. Pairing it with 3DGS for scene reconstruction tightens the Real2Sim loop in a way that could meaningfully reduce the manual labor cost of building training environments.

---

## Key Takeaways

- A six-author team has published an end-to-end VLA training and deployment pipeline running entirely on AMD ROCm hardware, covering data-center training, simulation rendering, and edge inference.
- Four demonstrations span sim-to-real manipulation (SmolVLA on Franka), language-grounded selection, Real2Sim data generation via 3D Gaussian Splatting and Genesis, and humanoid/quadruped locomotion RL across multiple hardware platforms.
- All pipelines run on RDNA4 (Radeon AI PRO R9700) and RDNA3.5 (Radeon PRO W7900) with ROCm and PyTorch, and are reproducible on the free Radeon Cloud Platform.
- The paper does not provide quantitative CUDA vs. ROCm throughput comparisons — a gap that limits its immediate utility for teams evaluating hardware switching costs.
- Strategically, this positions AMD's open ROCm stack as a functional alternative for Physical AI infrastructure, directly relevant to the compute supply chain that humanoid robotics companies depend on.
- The Real2Sim 3DGS pipeline for synthetic data generation may be the most immediately actionable contribution for labs looking to reduce simulation authoring costs.

---

## Frequently Asked Questions

**Can VLA models for humanoid robots be trained without NVIDIA GPUs?**
This paper demonstrates that VLA-based manipulation and locomotion policies can be trained and deployed on AMD GPUs using the ROCm software stack without CUDA. However, the paper does not publish head-to-head throughput benchmarks against NVIDIA hardware, so performance parity claims cannot be independently verified from the source material alone.

**What is Real2Sim2Real and why does it matter for humanoid robotics?**
Real2Sim2Real refers to the loop of scanning real environments into simulation (Real2Sim), training policies in simulation, then deploying back to physical robots (Sim-to-Real). For humanoid robotics, this approach reduces the cost of manually building simulation assets and can improve policy transfer by training on photorealistic reconstructions of actual deployment spaces.

**What is 3D Gaussian Splatting and how is it used here?**
3D Gaussian Splatting (3DGS) is a scene reconstruction technique that represents environments as collections of 3D Gaussian primitives, enabling fast, high-quality rendering. In this pipeline, 3DGS reconstructions of real scenes are fused with the Genesis physics engine to generate synthetic training data, tightening the Real2Sim pipeline without manual environment modeling.

**What is SmolVLA and why use it for this demonstration?**
SmolVLA is a compact, open-weight vision-language-action model. The paper uses it for the baseline sim-to-real manipulation demonstration on a Franka arm. The source material does not provide architectural specifics or parameter counts beyond the model name.

**Does AMD's ROCm stack support large-scale reinforcement learning for humanoid locomotion?**
According to the paper, large-scale RL for humanoid and quadruped locomotion was benchmarked across multiple hardware platforms using the ROCm stack. Specific benchmark scores or comparisons to CUDA-based training runs are not provided in the source material.