Hardware & Lab Architecture
Why Physical AI Needs Serious Compute
Physical AI systems are computationally demanding because they must:
- Process sensor data in real-time: Cameras, LiDAR, IMU generate massive data streams
- Run AI models at inference speed: Perception, planning, and control require low latency
- Handle multiple parallel tasks: Navigation, manipulation, and communication simultaneously
- Support sim-to-real development: Training models in simulation before deployment
This section describes the hardware requirements and lab setup options for Physical AI development.
Digital Twin Workstation Requirements
Your development workstation runs simulations, trains models, and develops code before deploying to physical robots.
Minimum Requirements
- CPU: 6+ cores (Intel i7/AMD Ryzen 7 or equivalent)
- RAM: 16GB (32GB recommended for large simulations)
- GPU: NVIDIA RTX 3060 or better (required for Isaac Sim)
- Storage: 100GB+ free space (SSD recommended for faster I/O)
- OS: Ubuntu 22.04 LTS
Recommended Configuration
- CPU: 12+ cores (Intel i9/AMD Ryzen 9)
- RAM: 32GB or 64GB
- GPU: NVIDIA RTX 4070 or better (RTX 4090 for best performance)
- Storage: 500GB+ NVMe SSD
- Network: Gigabit Ethernet for cloud integration
Why GPU Matters
- Isaac Sim: Requires RTX GPU for real-time ray tracing
- Model Training: GPU acceleration speeds up training by 10-100x
- Real-time Inference: Some models require GPU for acceptable latency
Physical AI Edge Kit
For deploying to real robots, you need edge computing devices that can run AI models with low latency.
NVIDIA Jetson Options
Jetson Orin Nano (Recommended):
- GPU: 1024-core NVIDIA Ampere architecture
- RAM: 8GB
- Power: 7-15W
- Best for: Small robots, cost-sensitive projects
Jetson Orin NX:
- GPU: 1024-core NVIDIA Ampere architecture
- RAM: 16GB
- Power: 10-25W
- Best for: Medium robots, more complex AI workloads
Jetson AGX Orin:
- GPU: 2048-core NVIDIA Ampere architecture
- RAM: 32GB or 64GB
- Power: 15-60W
- Best for: Large robots, research platforms
Sensor Suite
Depth Camera (Intel RealSense D435/D455):
- RGB-D sensing for navigation and manipulation
- ROS 2 support via
realsense-ros - ~$200-300
IMU (Bosch BMI160 or similar):
- Inertial measurement for odometry and stabilization
- Integrated into many robot platforms
- ~$20-50
Microphone Array (ReSpeaker or similar):
- Multi-microphone array for voice commands
- Beamforming for noise reduction
- ~$50-100
LiDAR (Optional, for advanced navigation):
- 2D or 3D LiDAR for mapping and localization
- More expensive ($500-2000+)
- Often integrated into robot platforms
Robot Lab Options
Tier 1: Proxy Robot
Options:
- Unitree Go2: Quadruped robot, ~$1,600
- Robotic Arm: 6-DOF arm (e.g., Dynamixel-based), ~$500-2000
- Mobile Base: Differential drive platform, ~$300-800
Pros:
- Lower cost entry point
- Good for learning ROS 2 and basic behaviors
- Less risk of damage during development
Cons:
- Limited manipulation capabilities (for non-arm options)
- May not represent humanoid challenges
- Less impressive for demonstrations
Tier 2: Miniature Humanoid
Options:
- Hiwonder XArm: Small humanoid, ~$300-500
- OP3 (Open Platform 3): Research humanoid, ~$1,000-1,500
- Unitree G1 Mini: Compact humanoid, ~$3,000-5,000
Pros:
- Humanoid form factor (bipedal, arms, head)
- Good for manipulation and navigation research
- More realistic for humanoid AI development
Cons:
- Higher cost than proxy robots
- May have limited payload and stability
- Requires more careful control
Tier 3: Premium Lab (Unitree G1)
Unitree G1 (Full-size humanoid):
- Cost: ~$16,000-20,000
- Capabilities: Full humanoid with advanced manipulation
- Best for: Research labs, sim-to-real transfer, publication-quality work
Pros:
- Industry-leading humanoid platform
- Excellent sim-to-real transfer capabilities
- Impressive demonstrations and research outcomes
Cons:
- High cost (may require grant funding)
- Requires significant space and safety considerations
- May be overkill for learning/teaching
Ether Lab (Cloud-Native) Setup
For teams or individuals without local GPU workstations, cloud-based development is an option.
Cloud Workstation Options
AWS EC2 Instances:
- g5.xlarge: 1x A10G GPU, 4 vCPU, 16GB RAM (~$1.00/hour)
- g5.2xlarge: 1x A10G GPU, 8 vCPU, 32GB RAM (~$1.50/hour)
- g6e.xlarge: 1x L4 GPU, 4 vCPU, 32GB RAM (~$0.75/hour)
Google Cloud Platform:
- n1-standard-4 + T4 GPU: ~$0.50-1.00/hour
- a2-highgpu-1g: 1x A100 GPU, ~$3.00/hour
Azure:
- NC6s v3: 1x V100 GPU, ~$2.00/hour
- NCas_T4_v3: 1x T4 GPU, ~$0.80/hour
Cost Considerations
OpEx (Operational Expenditure) - Cloud:
- Pay-as-you-go pricing
- No upfront hardware investment
- Easy to scale up/down
- Can become expensive with heavy usage
CapEx (Capital Expenditure) - Local:
- One-time hardware purchase
- No ongoing compute costs
- Better for long-term, intensive use
- Requires upfront investment
When to Use Cloud
- Limited budget: Can't afford local GPU workstation
- Team sharing: Multiple developers can share instances
- Occasional use: Don't need GPU 24/7
- Experimentation: Try different GPU types before buying
When to Use Local
- Intensive development: Using GPU many hours per day
- Data privacy: Sensitive robot data shouldn't leave premises
- Low latency needs: Local development is faster
- Long-term projects: Cost-effective over 6+ months
The Latency Trap
Why Cloud Control is Dangerous
Network Latency:
- Typical cloud latency: 50-200ms (depending on location)
- Robot control loops require: <10ms for safety
- Result: Unstable, unsafe robot behavior
Connection Reliability:
- Network drops cause robot to lose control
- Safety-critical systems cannot tolerate interruptions
- Real robots can cause physical harm if uncontrolled
The Solution: Train-in-Cloud, Deploy-to-Jetson
Development Phase (Cloud):
- Train AI models in cloud workstations (fast, powerful GPUs)
- Develop and test code in cloud simulations
- Iterate quickly without local hardware constraints
Deployment Phase (Edge):
- Deploy optimized models to Jetson edge devices
- Run inference locally on the robot (low latency, reliable)
- Use cloud only for monitoring and data collection
Pattern:
Cloud Workstation (Training)
↓ (export model)
Jetson Edge Device (Inference)
↓ (robot control)
Physical Robot
Lab Planning Guidance
For Instructors
Budget Considerations:
- Minimal: Cloud workstations + proxy robots (~$2,000-5,000 total)
- Standard: Local workstations + miniature humanoids (~$10,000-20,000)
- Premium: High-end workstations + Unitree G1 (~$30,000-50,000)
Student Access:
- Shared lab workstations for simulation
- Rotating access to physical robots
- Cloud backup for overflow capacity
For Self-Learners
Starting Point:
- Cloud workstation for initial learning (AWS free tier or low-cost instances)
- Proxy robot (Unitree Go2 or robotic arm) for hands-on experience
- Jetson Orin Nano for edge deployment
Progression:
- Start with simulation-only (Gazebo, Isaac Sim)
- Add proxy robot for real-world experience
- Upgrade to humanoid when ready for advanced projects
Budget Planning Worksheet
| Item | Tier 1 (Minimal) | Tier 2 (Standard) | Tier 3 (Premium) |
|---|---|---|---|
| Workstation | Cloud ($100-500/mo) | Local ($2,000-4,000) | Local ($5,000-8,000) |
| Edge Device | Jetson Orin Nano ($500) | Jetson Orin NX ($700) | Jetson AGX Orin ($2,000) |
| Sensors | RealSense ($300) | RealSense + IMU ($400) | Full suite ($1,000) |
| Robot | Proxy ($500-2,000) | Mini Humanoid ($1,000-5,000) | Unitree G1 ($16,000) |
| Total | $1,300-3,300 | $4,100-10,100 | $24,000-27,000 |