Automation

AWS, fiziksel AI geliştirme için açık kaynak araç zinciri başlattı

AWS, fiziksel AI geliştirme için açık kaynak araç zinciri başlattı

Key Takeaways

  • AWS + NVIDIA tam açık kaynak Physical AI Toolchain sunar; veri toplama, sentetik senaryo oluşturma, model eğitimi, simülasyon, doğrulama ve uç dağıtımını kapsar.
  • Yığın, üretim-düzeyi altyapı için Amazon SageMaker, EC2 p4d.24xlarge (8 × NVIDIA A100 GPU, 96 vCPU, 1.1 TB RAM), S3 ve IoT Greengrass kullanır.
  • NVIDIA bileşenleri Isaac Sim, Isaac Lab, Isaac GR00T ve Cosmos’u içerir; hepsi NVIDIA OSMO tarafından yönlendirilir.
  • Geliştiriciler herhangi bir robot tanımını (URDF, LeRobot) ve görev verisini bağlayabilir; bu sayede mühendislik süresinde %40’a kadar tasarruf sağlanır (ilk AWS müşterileri bildiriyor).
  • Araç zinciri, depo, enerji, sağlık, madencilik, tarım, havacılık ve savunma gibi yüksek etkili dikeylere yöneliktir.

Physical AI Toolchain Genel Bakışı

Amazon Web Services, AWS’in elastik bulutunu NVIDIA’nın robotik yazılım paketine birleştiren açık kaynak Physical AI Toolchain’i yayınladı. Teklif, hazır referans mimarileri, Infrastructure-as-Code (IaC) şablonları ve bir fizik-AI sisteminin tüm yaşam döngüsünü kapsayan otomatik dağıtım boru hatları (ham robot telemetrisi’den uç cihazda çalışan modele) sağlar.

“Müşteriler yenilik yerine altyapıya çok fazla zaman harcıyordu,” diyor Uwem Ukpong, AWS Industries VP’si. “Araç zinciri bu dengeyi değiştiriyor.”

Temel Tasarım İlkeleri

Principle How It’s Implemented
Modularity Individual components (e.g., Isaac Sim, SageMaker) can be used alone or chained together.
Hardware Agnostic Accepts custom robot descriptions (URDF, LeRobot) and task data, supporting manipulators, mobile bases, and humanoids.
End-to-End Automation IaC scripts provision EC2 p4d.24xlarge for training, S3 buckets for data, and Greengrass groups for edge rollout.
Open-Source All code, from orchestration scripts to sample notebooks, is publicly available on GitHub under the Apache 2.0 license.

Sentetik Veriden Uç Dağıtıma

Synthetic Data Generation

  • NVIDIA Cosmos creates photorealistic training scenes at up to 10 kHz frame rates, exporting data in ONNX and ROS 2 compatible formats.

Model Training

  • Amazon SageMaker (ml.p4d.24xlarge) provides 8 × A100 GPUs, delivering ~312 TFLOPS of FP16 compute for reinforcement-learning or supervised pipelines.
  • Native support for PyTorch, Hugging Face Transformers, and Gymnasium environments.

Simulation & Validation

  • Isaac Sim runs on the same EC2 GPU fleet, delivering real-time physics at 1 ms time steps for complex manipulators.
  • Isaac Lab and Isaac GR00T enable reinforcement-learning loops that converge up to 2× faster than on-prem clusters, thanks to NVIDIA’s TensorRT-accelerated inference.

Edge Deployment

  • Trained models are packaged as ONNX graphs and pushed via AWS IoT Greengrass to edge devices (e.g., NVIDIA Jetson AGX Orin).
  • Continuous improvement pipelines pull operational telemetry back into S3, triggering automated re-training cycles in SageMaker.

Comparison: Traditional In-House Pipeline vs. AWS Physical AI Toolchain

Aspect Traditional In-House Setup AWS Physical AI Toolchain
Compute Provisioning Fixed on-prem GPU servers (often 4 × RTX 3090, 48 vCPU, 384 GB RAM) On-demand EC2 p4d.24xlarge (8 × A100, 96 vCPU, 1.1 TB RAM)
Software Stack Manual integration of ROS, custom simulators, proprietary data pipelines Pre-integrated NVIDIA Isaac suite + AWS services (SageMaker, S3, Greengrass)
Scalability Limited by physical rack space; scaling takes weeks Elastic scaling within minutes; pay-as-you-go
Time-to-Market 6–12 months for full pipeline build 2–4 weeks using reference IaC templates
Cost Model CAPEX heavy, OPEX unpredictable OPEX transparent; typical training job ≈ $2,400 per 100 M steps on p4d.24xlarge
Maintenance Dedicated sysadmin team required Managed services handle patching, security, and backups

Real-World Example

AWS ships a sample workflow that trains a UR3 pick-and-place robot using 27 tele-operation episodes captured in the LeRobot format. The pipeline demonstrates:

  1. Data ingestion into S3 (≈ 2 GB).
  2. Synthetic augmentation via Cosmos (10× more scenarios).
  3. Reinforcement learning in Isaac GR00T on a single SageMaker training job (≈ 4 h).
  4. Simulation validation in Isaac Sim (real-time).
  5. Edge deployment to a Jetson Orin module via Greengrass.

The end-to-end run validates a 95 % success rate on unseen objects, a figure that rivals bespoke commercial solutions.

Target Industries

AWS highlights eight sectors where the toolchain can accelerate ROI:

  • Industrial Automation – robotic cell optimization.
  • Warehousing & Logistics – autonomous picking and sorting.
  • Energy – inspection drones and robotic manipulators.
  • Healthcare

İlgili Makaleler