Building the Physical AI Lab

Introduction: The Infrastructure Decision

Before you write your first line of robot code, you face a critical infrastructure decision: Cloud or Physical? This isn't just about cost—it's about latency, scalability, maintenance, and learning objectives.

The Hybrid Approach

Most successful Physical AI programs use both: cloud for training compute-intensive models, physical hardware for real-time inference and robot control.

Option 1: The "Ether Lab" (Cloud Infrastructure)

Why Cloud?

The cloud-based approach eliminates upfront capital expenditure (CapEx) and provides instant access to high-end GPUs without hardware maintenance.

Advantages:

✅ Zero upfront investment
✅ Instant provisioning (spin up instances in 2 minutes)
✅ Access to latest hardware (H100s, A100s on demand)
✅ No cooling, power, or space requirements
✅ Perfect for 10-week courses (no equipment storage)

Disadvantages:

❌ Ongoing operational costs (OpEx)
❌ Network latency for real-time control
❌ Cannot connect to physical robots directly
❌ Data egress fees (downloading models)
❌ Credit exhaustion risk for students

Recommended Cloud Configuration

AWS g5.2xlarge Instance

Specifications:

GPU: 1x NVIDIA A10G Tensor Core (24GB VRAM)
vCPUs: 8 cores (3rd Gen AMD EPYC)
RAM: 32 GB
Storage: 225 GB NVMe SSD
Network: Up to 10 Gbps
Cost: ~$1.21/hour (On-Demand, us-east-1)

Monthly Cost Estimate (8 hours/day, 20 days/month):

$1.21/hour × 8 hours/day × 20 days = $193.60/month

Cost Management

Enable auto-stop after 2 hours of inactivity to prevent accidental overnight charges ($30+ per night).

AWS Setup Guide

Step 1: Launch Instance

# Install AWS CLI
curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
unzip awscliv2.zip
sudo ./aws/install

# Configure credentials
aws configure

# Launch g5.2xlarge with Deep Learning AMI
aws ec2 run-instances \
  --image-id ami-0c7217cdde317cfec \
  --instance-type g5.2xlarge \
  --key-name your-keypair \
  --security-group-ids sg-xxxxxxxxx \
  --subnet-id subnet-xxxxxxxxx \
  --block-device-mappings '[{"DeviceName":"/dev/sda1","Ebs":{"VolumeSize":500}}]' \
  --tag-specifications 'ResourceType=instance,Tags=[{Key=Name,Value=PhysicalAI-Lab}]'

Step 2: Connect and Verify GPU

# SSH into instance
ssh -i your-keypair.pem ubuntu@<instance-ip>

# Verify GPU
nvidia-smi

# Expected output:
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 525.60.13    Driver Version: 525.60.13    CUDA Version: 12.0     |
# |-------------------------------+----------------------+----------------------+
# | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
# |   0  NVIDIA A10G         On   | 00000000:00:1E.0 Off |                    0 |

Step 3: Install Isaac Sim (Cloud)

# Install Omniverse Launcher
wget https://install.launcher.omniverse.nvidia.com/installers/omniverse-launcher-linux.AppImage
chmod +x omniverse-launcher-linux.AppImage

# Note: Isaac Sim headless mode for cloud
# Full GUI rendering requires X11 forwarding or VNC

Alternative Cloud Providers

Paperspace (More Affordable)

Configuration: P4000 instance

GPU: NVIDIA Quadro P4000 (8GB)
Cost: $0.51/hour ($81.60/month @ 8 hrs/day)
Best for: Students on tight budgets
Limitation: Cannot run Isaac Sim (needs 12GB+ VRAM)

Google Colab Pro+ (Quick Experiments)

GPU: V100 or A100 (varies)
Cost: $50/month (limited compute units)
Best for: Jupyter notebook experiments, not production
Limitation: 24-hour session timeout

Option 2: The "Sim Rig" (Physical Workstation)

Why Physical Hardware?

A dedicated workstation provides low-latency control for physical robots and avoids recurring cloud costs for long-term projects.

Advantages:

✅ One-time investment (own forever)
✅ Zero latency for robot control
✅ No data egress fees
✅ Can connect USB sensors (RealSense, ReSpeaker)
✅ Tax-deductible for educational institutions

Disadvantages:

❌ High upfront cost ($1,800 - $3,000)
❌ Requires physical space and cooling
❌ Maintenance responsibility
❌ Hardware obsolescence (3-5 year lifespan)

Recommended Workstation Configuration

The "Golden Build" ($1,849)

Component	Model	Specs	Price
GPU	NVIDIA RTX 4070 Ti	12GB GDDR6X, 7680 CUDA cores	$799
CPU	AMD Ryzen 9 7900X	12 cores, 24 threads, 4.7 GHz	$429
Motherboard	ASUS TUF Gaming X670E	PCIe 5.0, WiFi 6E	$289
RAM	Corsair Vengeance DDR5	32GB (2x16GB), 5600 MHz	$119
Storage	Samsung 980 Pro	500GB NVMe Gen4	$69
PSU	Corsair RM850x	850W 80+ Gold	$129
Case	Fractal Design Meshify 2	Airflow optimized	$159
Cooling	Noctua NH-D15	Dual tower, 140mm fans	$109

Total: $2,102 (before tax/shipping)

Why RTX 4070 Ti?

The 12GB VRAM is the minimum for NVIDIA Isaac Sim's recommended specs. RTX 4060 Ti (8GB) will struggle with complex scenes.

Alternative Budget Build ($1,299)

For students or hobbyists:

Component	Downgrade	Savings
GPU	RTX 4060 Ti (8GB)	-$300
CPU	Ryzen 7 7700X (8-core)	-$150
RAM	16GB DDR5 (2x8GB)	-$60
Motherboard	B650 chipset	-$100
Cooling	Corsair air cooler	-$50

New Total: $1,442

Trade-offs:

⚠️ Isaac Sim will run at low graphics settings
⚠️ Limited to 2-3 simulated robots simultaneously
⚠️ Cannot train large RL models locally

Operating System: Ubuntu 22.04 LTS

Why Ubuntu?

✅ ROS 2 Humble official support (until 2027)
✅ NVIDIA Isaac Sim native compatibility
✅ All tutorials assume Ubuntu/Linux
✅ Free and open-source

Installation Guide:

Download Ubuntu 22.04

# Download from official site
wget https://releases.ubuntu.com/22.04/ubuntu-22.04.3-desktop-amd64.iso

# Create bootable USB (on Linux)
sudo dd if=ubuntu-22.04.3-desktop-amd64.iso of=/dev/sdX bs=4M status=progress

Dual Boot Configuration (Optional)
- Keep Windows for gaming, Ubuntu for robotics
- Allocate 200GB+ partition for Ubuntu
- Install GRUB bootloader

Post-Installation: NVIDIA Drivers

# Check GPU detection
lspci | grep -i nvidia

# Install latest NVIDIA driver
sudo ubuntu-drivers devices
sudo ubuntu-drivers autoinstall

# Reboot and verify
sudo reboot
nvidia-smi

CapEx vs OpEx Analysis

5-Year Total Cost of Ownership

Cloud (AWS g5.2xlarge):

Cost per hour: $1.21
Hours per week: 40 (full-time usage)
Weeks per year: 50 (accounting for breaks)

Annual cost: $1.21 × 40 × 50 = $2,420/year
5-year cost: $2,420 × 5 = $12,100

Physical Workstation:

Initial investment: $1,849
Electricity: ~$15/month × 12 × 5 = $900
Upgrades (RAM, storage): $300 over 5 years

5-year cost: $1,849 + $900 + $300 = $3,049

Break-even point: ~18 months of regular usage

Institutional Decision

Universities: Physical workstations (multi-student access, 10+ year amortization)
Bootcamps/Short Courses: Cloud (no maintenance, pay-per-use)
Individual Learners: Physical if committed to 2+ years, Cloud for experimentation

Recommended Lab Layout

Small-Scale Lab (1-10 Students)

Configuration:

5x Physical Workstations (RTX 4070 Ti)
1x High-end Server (Dual RTX 4090s for shared training)
1x NAS (Network Attached Storage) for datasets
10Gbps local network

Cost: ~$15,000 initial + $200/month electricity

Large-Scale Lab (50+ Students)

Configuration:

2x Physical Workstations (instructor demos)
AWS Cloud Credits ($100/student/semester)
Remote desktop infrastructure (Apache Guacamole)
GitLab instance for code collaboration

Cost: ~$10,000 setup + $5,000/semester cloud

Space and Power Requirements

Physical Requirements

Per Workstation:

Space: 2' × 2' desk footprint
Power: 850W PSU (draws 300-600W typical)
Cooling: Room AC recommended (generates ~1,200 BTU/hour)
Ventilation: Avoid enclosed cabinets (GPU needs airflow)

Classroom Setup:

Electrical: 15A circuit per 2 workstations
Cooling: 1 ton AC per 5 workstations
Fire Safety: Ensure proper extinguisher access

Fire Hazard

Never run more than 3 high-power GPUs on a single 15A household circuit. Use dedicated 20A circuits for labs.

Maintenance and Upkeep

Physical Workstation Maintenance

Monthly:

Clean dust filters (every 30 days in dusty environments)
Check GPU temperatures (below 80°C under load)
Backup important datasets

Quarterly:

Update NVIDIA drivers
Update Ubuntu kernel
Repaste thermal compound if temps exceed 85°C

Annual:

Replace case fans (if noisy or failing)
Upgrade storage if under 20% free space
Consider GPU upgrade if under 8GB VRAM

Decision Framework

Choose Cloud if:

✅ Course duration under 6 months
✅ Budget is flexible (OpEx model approved)
✅ No physical robot hardware
✅ Students work remotely
✅ Need access to H100/A100 GPUs

Choose Physical if:

✅ Long-term commitment (2+ years)
✅ Budget has upfront capital available
✅ Physical robots in lab (Unitree, Boston Dynamics)
✅ Low-latency control required
✅ Want to teach hardware maintenance

Hybrid Approach (Recommended)

Best of Both Worlds

Local Workstations (1-2 units):

Robot control and testing
Real-time inference (Jetson connected)
Student project demonstrations

Cloud Instances (AWS/GCP):

Large-scale RL training (Isaac Gym)
Synthetic data generation (10,000+ images)
Model fine-tuning (LoRA, RLHF)

Workflow:

graph LR
    A[Local Workstation] -->|Upload Model| B[Cloud GPU]
    B -->|Train 24 hours| C[Trained Weights]
    C -->|Download| A
    A -->|Deploy| D[Jetson Orin Nano]
    D -->|Control| E[Physical Robot]

Troubleshooting Common Issues

Cloud Problems

Issue: "Insufficient capacity" error on AWS

# Solution: Try different availability zone
aws ec2 run-instances ... --availability-zone us-east-1b

Issue: Slow SSH connection

# Solution: Enable connection multiplexing
# Add to ~/.ssh/config:
Host *.compute.amazonaws.com
    ControlMaster auto
    ControlPath ~/.ssh/sockets/%r@%h-%p
    ControlPersist 600

Physical Hardware Problems

Issue: GPU not detected in Ubuntu

# Solution: Secure boot may block NVIDIA drivers
# Disable in BIOS or sign drivers:
sudo mokutil --disable-validation

Issue: Overheating (greater than 85°C)

# Check fan speeds
nvidia-smi -q -d TEMPERATURE

# Solution: Improve case airflow or repaste GPU

Summary

Aspect	Cloud (AWS g5.2xlarge)	Physical (RTX 4070 Ti)
Upfront Cost	$0	$1,849
Monthly Cost	$194 (8hr/day)	$15 (electricity)
5-Year Cost	$11,640	$2,749
GPU VRAM	24GB (A10G)	12GB
Robot Control	❌ High latency	✅ Real-time
Maintenance	Zero	DIY required
Best For	Short courses	Long-term projects

Next Steps

Evaluate Budget: Calculate CapEx vs OpEx for your timeline
Assess Use Case: Do you need physical robot integration?
Prototype First: Start with cloud, migrate to physical if committed
Plan Space: Ensure adequate power and cooling for physical rigs

Recommended Path for Most Learners:

Start with AWS free tier (limited)
Use Paperspace for experimentation ($50-100/month)
Invest in physical workstation if continuing beyond 6 months

Pro Tip

Many universities have AWS Educate or Azure for Students credits ($100-200/year). Check with your institution before paying out of pocket.

Next Chapter: The Student Kit (Jetson & Sensors)

Introduction: The Infrastructure Decision​

Option 1: The "Ether Lab" (Cloud Infrastructure)​

Why Cloud?​

Recommended Cloud Configuration​

AWS g5.2xlarge Instance​

AWS Setup Guide​

Alternative Cloud Providers​

Paperspace (More Affordable)​

Google Colab Pro+ (Quick Experiments)​

Option 2: The "Sim Rig" (Physical Workstation)​

Why Physical Hardware?​

Recommended Workstation Configuration​

The "Golden Build" ($1,849)​

Alternative Budget Build ($1,299)​

Operating System: Ubuntu 22.04 LTS​

CapEx vs OpEx Analysis​

5-Year Total Cost of Ownership​

Recommended Lab Layout​

Small-Scale Lab (1-10 Students)​

Large-Scale Lab (50+ Students)​

Space and Power Requirements​

Physical Requirements​

Maintenance and Upkeep​

Physical Workstation Maintenance​

Decision Framework​

Choose Cloud if:​

Choose Physical if:​

Hybrid Approach (Recommended)​

Best of Both Worlds​

Troubleshooting Common Issues​

Cloud Problems​

Physical Hardware Problems​

Summary​

Next Steps​