Physical AI · Plan

Physical AI Build Plan

This is the sequenced hardware purchase and project plan that follows the Learning Plan. It runs in three hardware tiers, from a ~$1,000 starter kit up to humanoid-grade compute. The six projects build on each other, from Jetson hello-world to robotic-arm imitation learning to Claude-to-robot via AgenticROS. From there: multi-camera vision, offline voice assistant, autonomous navigation, and sim-to-real exploration in Isaac Sim. The principle is simple. Hardware purchases serve learning goals, projects build skills, and the Learning Plan informs the decisions.

Plan   Last updated May 17, 2026

Context

This plan captures a sensible starter path for Physical AI builders based on May 2026 hardware availability. All items are candidates, not fixed prescriptions: tier and project order should be informed by your specific learning goals from the Learning Plan.

The principle: every hardware purchase should serve a learning goal, not just be "cool to have." Each project should build a skill that maps to where you want to go. Spending time in Learning Plan Phase 1 through 3 before locking Tier 1 hardware is the right move.

Hardware tiers

Tier 1: starter (~$1,000)

The broadest learning surface for the lowest investment and wait time.

ItemPriceShipsLearning goal
NVIDIA Jetson Orin Nano Super Dev Kit$249ImmediatelyEdge AI deployment, CUDA, JetPack, local LLM on dedicated hardware, computer vision, foundation for all other projects
SO-ARM101 (assembled, from PartaBot)~$300ImmediatelyImitation learning, LeRobot framework, PyTorch on real hardware
Reachy Mini Wireless$449~30 days (was 90+; ramped up)Hugging Face robotics ecosystem, expressive robot interaction, multi-modal AI (vision + speech + motion). App Store launched May 6, 2026 with 200+ apps.

Total Tier 1: ~$1,000.

Tier 2: expansion

Expand the learning surface after Tier 1 projects are running.

ItemPriceLearning goal
Intel RealSense D435 depth camera~$3143D spatial perception (depth, not just 2D images). Standard sensor for robotic arms and navigation. Used in AgenticROS setups.
USB cameras (2-3 for multi-camera vision)~$50-100Computer vision, DeepStream multi-camera pipelines, home security project
USB microphone + speaker for voice assistant~$50Whisper STT + local LLM + TTS pipeline on Jetson
JetBot or JetRacer kit$200-500Autonomous navigation, reinforcement learning, SLAM

Tier 3: aspirational

Not for immediate purchase. Revisit after completing Tier 1 and Tier 2.

ItemPriceLearning goal
Unitree Go2 quadruped~$2,800Quadruped locomotion AI, LIDAR, autonomous navigation at scale.
Jetson Thor Dev Kit$1,999+Foundation-model deployment, humanoid-grade compute. Only if direction confirms need for Thor-class development.
Unitree G1$16,000+ baseFull humanoid robotics, GR00T N1.7, walking + manipulation. The reference embodiment for GR00T whole-body control. Major investment.
Unitree H2$29,900 commercial / $40,900 EDULarger humanoid (182 cm, 70 kg, 31 DOF). Shipping April 2026.

Projects (sequenced)

Each project maps to a Learning Plan phase and builds on the previous project.

Project 1: Jetson hello-world

Prerequisite: Jetson Orin Nano Super purchased and set up.

Skills built: JetPack SDK, edge AI deployment, local LLM inference, basic computer vision.

  1. Flash JetPack 6.2, set up the board, connect display and keyboard.
  2. Run a local LLM (Llama 8B or similar) and benchmark inference speed.
  3. Connect a USB camera and run real-time object detection with YOLO26.
  4. Set up DeepStream for a simple single-camera analytics pipeline.
  5. Run Whisper Large-v3 Turbo for speech-to-text and compare to cloud APIs.

Success criteria: the Jetson runs independently, performs inference, processes camera input in real time, and you understand the JetPack environment.

Project 2: Robotic arm + imitation learning

Prerequisite: Project 1 complete; SO-ARM101 purchased and assembled.

Skills built: imitation learning, LeRobot framework, PyTorch training loop, motor control, sim-to-real basics.

  1. Assemble and calibrate the SO-ARM101 (leader + follower).
  2. Follow the LeRobot tutorial: record demonstrations with the leader arm.
  3. Train a policy from the demonstration data.
  4. Deploy the policy to the follower arm and evaluate performance.
  5. Iterate: more demonstrations, better policy, harder tasks.

Success criteria: the follower arm reproduces a simple task (pick and place an object) learned from your demonstrations.

Project 2b: AgenticROS · connect Claude to your robot

Prerequisite: Project 2 complete (working arm).

Skills built: AgenticROS, ROS2 fundamentals, MCP in a physical context, natural-language robot control, NemoClaw governance layer.

  1. Install AgenticROS on the Jetson (follow the quick start at agenticros.com).
  2. Connect Claude Desktop or Claude Code to the arm via AgenticROS.
  3. Control the arm with natural-language commands through Claude.
  4. Add a RealSense D435 depth camera for spatial perception (if purchased).
  5. Explore NemoClaw for safety guardrails on the agent's actions.
  6. Document the Claude-to-robot pipeline (portfolio content with a governance angle).

Success criteria: you can type "pick up the red block" in Claude and the arm executes it. You understand how AgenticROS bridges AI agents to ROS2, and how NemoClaw monitors the agent's intent.

Why this project is special: it connects software-AI skills (Claude, MCP) to Physical AI skills (Jetson, ROS2, robotics). The lowest-friction path from software AI to Physical AI. It connects directly to NemoClaw governance and the broader intent-inspection direction.

Project 3: Multi-camera vision system

Prerequisite: Project 1 complete; USB cameras purchased.

Skills built: DeepStream multi-camera pipelines, real-time detection, alert systems, edge AI at scale.

  1. Set up 2-3 cameras on the Jetson Orin Nano.
  2. Build a DeepStream pipeline that processes all streams simultaneously.
  3. Detect people, packages, animals, vehicles.
  4. Set up Telegram or email alerts on detection events.
  5. Run 24/7 as a practical home system.

Success criteria: a functioning multi-camera AI system running locally, no cloud, with real-time alerts.

Project 4: Offline voice assistant

Prerequisite: Project 1 complete; mic + speaker purchased.

Skills built: Whisper STT, TTS (Piper / Kokoro), wake-word detection, LLM integration, multi-modal pipeline.

  1. Set up Whisper on Jetson for real-time speech-to-text.
  2. Connect to a local LLM on Jetson.
  3. Add text-to-speech output.
  4. Add wake-word detection (always-on listening, low power).
  5. Build a complete conversational loop.

Success criteria: say "Hey Jetson" (or custom wake word), ask a question, get a spoken answer, all running locally with zero cloud dependency.

Project 5: Autonomous navigation (JetBot / JetRacer)

Prerequisite: Projects 1-2 complete; navigation hardware purchased.

Skills built: reinforcement learning, SLAM, path planning, sensor fusion (camera + optional depth).

  1. Build the JetBot or JetRacer kit.
  2. Follow the NVIDIA JetBot tutorials for basic object following.
  3. Train autonomous navigation via RL (the robot learns to navigate a room).
  4. Add obstacle avoidance.
  5. Experiment with SLAM mapping.

Success criteria: the vehicle autonomously navigates a room, avoids obstacles, and returns to a starting position.

Project 6: Isaac Sim exploration

Prerequisite: Project 2 (arm experience) helps contextualize simulation.

Skills built: simulation-based training, digital twins, synthetic data generation, sim-to-real pipeline.

  1. Install Isaac Sim (workstation or cloud instance).
  2. Load a sample robotic arm environment.
  3. Generate synthetic training data.
  4. Train a policy in simulation.
  5. Attempt deploying the sim-trained policy to the real SO-ARM101.

Success criteria: you have completed a basic sim-to-real transfer. A skill learned in Isaac Sim works on your physical arm.

What is NOT in this plan (yet)

Explicitly deferred until the foundation is solid:

  • GR00T fine-tuning: requires deeper understanding of VLA models and likely Thor-class hardware.
  • Full humanoid projects: requires Unitree G1 or similar ($16K+) and extensive RL experience.
  • Defense-specific projects: separate exploration, not build projects.
  • Contributing to open-source robotics: earn credibility through building first.
  • Publishing about Physical AI: downstream of having learned something worth saying.

Cross-references

  • Learning Plan · the read-first study plan that informs hardware decisions.
  • Hardware Guide · sensors / compute / actuators / structure with verified prices.
  • Software Guide · runtime, perception, decision, world models, frameworks.
  • Systems Guide · what gets built across six form factors.
  • Governance Guide · the rules and the governance gaps NemoClaw begins to address.