Edge-to-Cloud Robotics with AMD ROCm: From Data Collection to Real-Time Inference

2026-03-23 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Cloud Computing & IT Infrastructure · Depth: Intermediate, long

Summary

AMD has developed an end-to-end edge-to-cloud robotics AI solution leveraging its hardware ecosystem and the Hugging Face LeRobot framework. This solution utilizes Imitation Learning (IL) and Vision-Language-Action (VLA) models like Pi0.5, Pi0, and smolVLA to teach robots complex tasks from demonstrations. The workflow involves collecting data on AMD Ryzen AI PCs, fine-tuning VLA models on high-performance AMD Instinct MI300X servers using the AMD ROCm software stack, and then deploying these trained models back to Ryzen AI PCs for real-time inference. This pipeline was showcased at AMD's 2025 Open Robotics Hackathons, where participants applied it to diverse scenarios such as food packaging, childcare, allergy aid, and waste classification, demonstrating its scalability and performance.

Key takeaway

For AI Engineers developing robotics solutions, this AMD-LeRobot pipeline offers a robust framework for rapid prototyping and deployment. You can efficiently train complex robotic behaviors using Imitation Learning with as few as 50 demonstration episodes, leveraging AMD's integrated hardware and software stack from edge data collection to cloud-based fine-tuning and real-time inference. Consider exploring the LeRobot GitHub and the AMD developer program for free cloud credits to accelerate your projects.

Key insights

AMD's edge-to-cloud robotics solution integrates LeRobot, VLA models, and Imitation Learning on AMD hardware for efficient robot training and deployment.

Principles

• Imitation Learning enables robots to learn from expert demonstrations.
• VLA models provide robots with rich visual and language understanding.
• Data-efficient foundation models require minimal examples for specialization.

Method

The method involves collecting visual and joint position data from a leader arm, fine-tuning pre-trained VLA models on cloud GPUs, and deploying the specialized policy to an edge PC for real-time inference.

In practice

• Use `lerobot-find-port` to identify SO-101 arm UART device nodes.
• Install PyTorch 2.7.1 with ROCm 6.3 for LeRobot compatibility.
• Set `HSA_OVERRIDE_GFX_VERSION=11.0.0` for Ryzen AI 300 series iGPUs.

Topics

• Edge-to-Cloud Robotics
• Imitation Learning
• Vision-Language-Action Models
• AMD ROCm
• Hugging Face LeRobot

Code references

• huggingface/lerobot
• TheRobotStudio/SO-ARM100
• huggingface/lerobot
• ROCm/AMD_Hackathon

Best for: AI Engineer, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Utilizing AMD Schola and UnrealRoboticsLab with AMD ROCm™ Software to Train a Robotic Arm — Integrating Unreal Engine, MuJoCo via URLab, and Schola enables high-fidelity, physics-accurate robotic arm reinforcement learning on AMD hardware.
• Train AI Robots Without Writing Code! (Introducing LeLab) — Leloup simplifies robot AI training via a GUI, enabling teleoperation, data collection, and model deployment without coding.
• Action Draft and Verify: A Self-Verifying Framework for Vision-Language-Action Model — ADV combines diffusion's precision with auto-regression's robustness for VLA models via a draft-and-verify mechanism.
• Physical Intelligence Team Unveils MEM for Robots: A Multi-Scale Memory System Giving Gemma 3-4B VLAs 15-Minute Context for Complex Tasks — MEM provides VLA models with dual-track memory for long-horizon robotic tasks, enabling adaptation and real-time performance.
• Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics — Ambient Diffusion Policy leverages noise-dependent data usage to extract useful features from suboptimal robotic demonstration data.
• SCOPE: Real-Time Natural Language Camera Agent at the Edge — SCOPE demonstrates real-time, natural-language PTZ camera control at the edge using optimized SLM/VLM combinations and quantization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.