Hi everyone,
While this community has mostly addressed questions from Developers, we also want to hear the great work you are doing with NVIDIA AI Aerial tools.
So, please share your highlights:
Let’s hear from you!
Many thanks to the Aerial team for your great work.
Since we do not currently have access to GH200, we have been waiting for a Spark-based version of the Aerial framework, and are looking forward to the cubb-26-1 release.
We also hope that ARC-OTA can support more O-RU vendors, for example Liteon FR1.
In addition, it would be great to see more AI-related demos or use cases so that we can better experience the performance and potential of AI-RAN.
Finally, we would really appreciate more GUI-based demos with richer visualizations to showcase the system in a more impactful and engaging way.
@xudong.zhao
yes, Liteon FR1 should work with Aerial-testbed (ARC-OTA).
Liteon FR2 has been validated in our lab, LiteON FR1 has been successfully integrated in other researcher’s lab with ARC-OTA1.7. We expect it works with Aerial-testbed 1.0 ( /ARC-OTA1.8, it is renamed to Aerial-testbed 1.0) running on dgx-spark.
Thanks,
This is really exciting to hear!
Regarding Liteon FR2 support
It’s truly great news that Liteon FR2 has already been validated and can be integrated. We noticed that at MWC 2026, synaXG demonstrated a setup using Spark + Liteon FR2.
We are very interested in the performance of that demo:
Is it capable of running full bandwidth (e.g., 400 MHz in FR2)?
Will there be documentation or tutorials in Aerial-testbed 1.0 explaining how to deploy and use this FR2 setup?
Hardware support roadmap
Besides NVIDIA DGX Spark and NVIDIA GH200, are there plans to support additional platforms such as ARC-Compact or other more lightweight / cost-efficient systems?
Release cadence
What is the expected update frequency for Aerial-testbed going forward?
We are very excited about the platform and would love to see new features and improvements delivered at a faster pace.
Roadmap visibility
Would it be possible to include a dedicated section in the Aerial-testbed documentation outlining the future roadmap?
This would greatly help the community align with NVIDIA’s direction and plan long-term development accordingly.
Thanks again for the great work — we are really looking forward to Aerial-testbed 1.0!
@jixu We appreciate the confirmation regarding Liteon FR2 and FR1. It’s very encouraging to see the validation progress.
We have also noticed the 26-1 release update, and we are now looking forward to the upcoming Aerial-testbed 1.0 (ARC-OTA 1.8) on DGX Spark.
Thanks again for your support.
We recently completed a validation of a deterministic identity and continuity control-plane layer (ACELOGIC™), tested under AI-RAN-aligned simulation conditions.
The system models distributed execution environments similar to NVIDIA Aerial concepts, with a focus on enforcing identity, authority, and lifecycle continuity before execution in autonomous systems.
Key validation results:
• Deterministic identity generation (SHA-256 fingerprint at agent birth)
• 100% duplicate prevention (50/50 blocked)
• 0 split-brain events across all scenarios
• Identity preserved across restart and migration (0 drift)
• SAFE_MODE enforcement during partition (no automatic recovery)
• Multi-verifier authority model with 100% agreement
• ~15 ms bounded control-plane overhead
The simulation models lifecycle transitions including restart, migration, partition, and adversarial conditions using a discrete-event approach aligned with AI-RAN architectural patterns.
This work explores what appears to be a missing layer in autonomous systems infrastructure:
→ deterministic identity
→ enforceable authority
→ lifecycle continuity across distributed environments
Open to connecting with others working on similar control-plane challenges in distributed AI systems.
Hi ,
I’ve been working on an AI-native PHY receiver prototype using NVIDIA AI Aerial (Sionna), focusing on hybrid model-driven + learning-based architectures for 6G.
Project Repository: https://github.com/NithishJ29/ai-ofdm-6g-
Key Outcomes
Built an end-to-end link-level pipeline:
OFDM → Rayleigh fading → LS channel estimation → MMSE equalization → LDPC (5G) → Demapping
Developed a neural demapper that leverages channel state (CSI) and noise variance as input features
Benchmarked Classical vs AI receiver chains under imperfect CSI
Observed consistent BER improvements in low–mid SNR regions using the AI-based demapper
Designed a modular simulation framework for rapid experimentation with PHY components
Technical Insights
Explicit inclusion of CSI (h) and noise variance significantly improves neural demapper performance
Multi-SNR training is critical for generalization
AI components can partially compensate for channel estimation errors, especially with LS estimators
Even lightweight models show gains → supports feasibility of low-complexity AI in RAN pipelines
Challenges / Gotchas
Handling Sionna API/version differences (channel models, OFDM abstractions)
Training sensitivity to input normalization and scaling
Trade-offs between simulation realism (pilots, interpolation) and training efficiency
Opportunities / Wishlist
Fully end-to-end differentiable PHY pipelines (channel → decoder)
Reference implementations for neural receivers (e.g., DeepRx-style)
Better tooling for dataset generation across diverse channel conditions
Pathways for real-time / hardware-in-loop validation
Closing Note
This project reinforces the direction of AI-native 6G PHY design, where learning-based components enhance—not replace—classical signal processing.
Looking forward to extending this into adaptive, intent-driven PHY systems aligned with AI-RAN evolution.
Thank you everyone for the comments. Please keep them coming!