Volume IV Physical AI Studio: Staff Implementation Master Plan

Comprehensive 13-Week Execution Roadmap (September 21 – December 18, 2026)

Target Audience: Postdoctoral Researcher (Andrea, ETH Zurich) & Course Instruction Staff Course Lead: Prof. Vijay Janapa Reddi Target Semester: Spring 2027 Studio Launch Target Platform: Arduino UNO Q (“Unikue” Qualcomm QRB2210 Linux + STM32U585 MCU) · Seeed Studio SO-101 Arm · Hugging Face LeRobot · SmolVLA & ACT Core Reference Architecture: Master Landing Page · Course Syllabus · Pre-Flight Guide · 12-Competency Matrix

---

1. Executive Summary & Operational Mandate

This master plan governs the 13-week physical implementation, hardware bring-up, and pedagogical qualification period between Monday, September 21, 2026, and Friday, December 18, 2026.

NoteThe “Student Zero” Methodology

To ensure every student station operates flawlessly when the studio launches in Spring 2027, the instruction team acts as Student Zero. Before any student touches the bench, staff personally unboxes, mounts, wires, programs, and qualifies every single component, lab brief, and automated test.

ImportantStudio Focus: The Dual-Brain Architecture

This studio teaches students to build and evaluate physical AI systems across two distinct computing layers: 1. The Cognitive Brain (Qualcomm QRB2210 Linux MPU): High-capacity neural policies (SmolVLA, ACT) that ingest sensory camera streams (\(I_t\)) and propose multi-step action trajectory chunks (\(a_{\text{req}}\)). 2. The Safety Governor (STM32U585 Real-Time MCU): Deterministic safety invariants (velocity saturation clamping \(\omega_i \le 45^\circ/\text{s}\), geometric table geofencing \(z_{\text{tool}} \ge 15\text{ mm}\), and \(150\text{ ms}\) communication watchdogs) that arbitrate, permit, or veto proposals (\(a_{\text{enf}}\)) before they reach physical motor coils (\(a_{\text{meas}}\)).

Students build, measure, and defend the Causal Permission Boundary connecting the two processors over an internal RPC bridge with zero unmonitored host bypass.

The Dual-Brain Architecture of Physical AI: Asymmetric Cognition vs. Deterministic Governance

The primary objectives for this 13-week execution window: 1. Week 1 Planning & Bench Preparation: Finalize the master curriculum plan, audit the Bill of Materials (BOM), allocate dedicated workbench space at ETH Zurich, and verify electrical power rails before unboxing. 2. Hardware Bring-Up & Validation: Assemble, wire, verify, and qualify one complete Golden Reference Station, validating the Four Hardware Validation Steps (Steps 1–4). 3. End-to-End Curriculum Qualification: Execute all 8 Hands-On Student Labs from scratch, producing verified golden artifacts, starter scripts, reference datasets, and calibration profiles. 4. Software & Firmware Freeze: Compile and lock the Golden Qualcomm Linux Image and STM32 Firmware Binary. 5. Independent Replication Audit: Have a secondary researcher or teaching assistant build Station 2 from scratch using only the staff documentation to ensure seamless class deployment.

---

2. Master 13-Week Execution Roadmap

The 13-week schedule is organized into six logical phases across five sprints, shifting physical assembly into Week 2 to allow Week 1 to focus entirely on architecture alignment, BOM audit, and bench preparation:

Volume IV Physical AI Studio: 13-Week Staff Implementation Roadmap

---

3. Physical Hardware Stack & Bench Architecture

The studio relies on an open, modular, and safety-hardened hardware stack designed specifically for edge Physical AI.

Physical AI Hardware Platform Overview

Component Image	System Pillar & Hardware	Key Technical Specifications	Core Function in Studio
	Dual-Silicon Compute Arduino UNO Q	• Qualcomm QRB2210 Linux Application MPU (Debian) • STM32U585 Real-Time Microcontroller (160 MHz ARM Cortex-M33) • OpenAMP RPMsg shared-SRAM inter-core bridge • Dedicated 45W USB-PD logic power	Runs Hugging Face LeRobot and SmolVLA/ACT neural policies on Linux; delegates physical motor bounds and safety governance to STM32.
	Manipulator Plant Seeed Studio SO-101 Pro	• 6-DoF open-source follower arm linkage • Rigid benchtop clamp mount (zero tip under load) • Parallel-jaw gripper end-effector • \(300 \times 200\text{ mm}\) marked manipulation workspace	Provides the physical body for imitation learning, human teleoperation demonstration recording, and autonomous block reach.
	Smart Serial Actuator Feetech STS3215 Bus Servo	• 19 kg·cm stall torque at 7.4V • 12-bit contactless magnetic angle encoder • 1 Mbps half-duplex UART daisy-chain protocol • Real-time angular position and load telemetry readback	Translates commanded joint action chunks (\(a_{\text{enf}}\)) into calibrated physical motion and returns measured angles (\(a_{\text{meas}}\)).
	Vision Sensor Logitech C270 HD Webcam	• 720p 30 Hz RGB video ingestion via Linux V4L2 • Mounted 40 cm overhead at 45° oblique perspective • AprilTag 36h11 extrinsic calibration target • Locked exposure and white balance to avoid visual drift	Ingests sensory frames \(o_t \in \mathbb{R}^{3 \times 224 \times 224}\) for closed-loop visual feedback and dynamic disturbance recovery.

Complete Bench Rigging & Electrical Bus Interface

Before assembling the physical arm or connecting motor power, review the electrical wiring schematic. The architecture uses clean galvanic separation between digital computing logic and high-current actuator power:

Physical Bench Rigging & Electrical Bus Interface

NoteDual Power Supply Engineering Rationale

To ensure bulletproof reliability and avoid mysterious CPU brownout resets: * Digital Logic Power: The Arduino UNO Q and USB camera are powered exclusively through the USB-C port via the 45W USB-PD adapter. * Actuator Power: The 6× STS3215 smart servos are powered by an independent, regulated 7.4V / 5A DC bench power supply connected directly to the servo bus rail. * Common Ground Reference: The DC motor power supply ground and Arduino UNO Q ground are connected together into a solid star-ground reference. Motor current NEVER passes through board headers. * Safety Invariant: Software faults, servo stalls, or power cuts on the motor rail will never reset the Qualcomm Linux processor or disrupt telemetry logging.

---

4. Sprint-by-Sprint Weekly Implementation Breakdown

Sprint 0: Architecture Alignment & Bench Preparation (Week 1)

Focus: Curriculum plan finalization, Bill of Materials audit, dedicated bench allocation, and electrical power verification.

Week 1 (Sep 21 – Sep 25, 2026): Master Plan Alignment & Bench Preparation

• Primary Objective: Finalize the master execution plan, complete the BOM inventory, allocate the physical workbench, and verify laboratory power systems.
• Tasks:
  • Review master curriculum architecture: Master Landing Page, Course Syllabus, Pre-Flight Guide, 8 Lab Briefs, and Capstone Studio.
  • Inventory delivered hardware packages: Seeed Studio SO-101 Pro arm kit, Arduino UNO Q (4GB RAM), 45W USB-PD power supply, USB 3.0 powered hub, Logitech C270 camera, 7.4V/5A DC motor power supply, UART level-shifter, and wiring harnesses.
  • Dedicate a clean \(120 \times 80\text{ cm}\) workbench at ETH Zurich with rigid table-edge clamping surface, overhead camera mount fixture, and anti-static mat.
  • Bench-test the regulated 7.4V DC motor supply and 45W USB-PD adapter with a digital multimeter to confirm clean voltage levels and common star-ground reference.
  • Stage bench workstation software: install Hugging Face LeRobot dependencies, PyTorch ARM64 toolchains, Arduino App Lab, and ONNX Runtime developer packages.
• Weekly Deliverable: Finalized master checklist, verified hardware inventory audit, workbench setup ready for assembly, and electrical schematic verification.

---

Sprint 1: Mechanical Bring-Up & Validation Steps 1–3 (Weeks 2–4)

Focus: Mechanical assembly, electrical isolation, and establishing the governed servo bus path.

Week 2 (Sep 28 – Oct 2, 2026): Hardware Bring-Up, Arm Assembly & Step 1 (USB Teleop)

• Primary Objective: Assemble the physical arm, establish power safety isolation, and achieve native teleoperation.
• Tasks:
  • Assemble Seeed Studio SO-101 6-DoF follower arm from kit components.
  • Clamp arm baseplate rigidly to workbench using heavy-duty C-clamps.
  • Wire external 7.4V/5A DC motor power supply directly to STS3215 servo power rail with common ground to UNO Q.
  • Connect SO-101 arm to workstation via USB BusLinker; install Hugging Face LeRobot (pip install lerobot).
  • Validation Step 1 Check: Run joint calibration and execute 60-second teleoperation replay (lerobot-replay).
• Weekly Deliverable: Video proof of Step 1 teleoperation + voltage rail multimeter verification.

Week 3 (Oct 5 – Oct 9, 2026): Inter-Core Bridge Wiring & Step 2 (1-Joint Interceptor)

• Primary Objective: Route servo communication through the STM32 and prove real-time command veto.
• Tasks:
  • Disconnect Joint 1 (Base Yaw) from USB BusLinker; connect single-wire half-duplex UART line to STM32U585 USART via level-shifter.
  • Flash baseline interceptor firmware to STM32 using Arduino App Lab.
  • Write Python bridge script on Qualcomm Linux transmitting target commands across /dev/ttyRPMSG.
  • Validation Step 2 Check: Verify STM32 permits valid commands (\(10^\circ\) at \(20^\circ/\text{s}\)), clamps over-speed commands (\(> 45^\circ/\text{s}\)), and refuses out-of-range targets (\(> 180^\circ\)).
• Weekly Deliverable: Telemetry log showing MCU clamping and rejection of out-of-bounds joint commands.

Week 4 (Oct 12 – Oct 16, 2026): Full 6-DoF Governed Arm & Step 3 (Full Integration)

• Primary Objective: Scale the governed bus to all 6 joints and achieve full LeRobot teleoperation under MCU control.
• Tasks:
  • Daisy-chain all 6 STS3215 servos into the STM32 USART bus line.
  • Implement the 60-line Python UnoQMotorsBus custom LeRobot adapter.
  • Program Cartesian forward kinematics check in STM32 firmware; implement table collision geofence (\(z_{\text{tool}} \ge 15\text{ mm}\)).
  • Validation Step 3 Check: Run full 6-DoF LeRobot teleoperation through the governed adapter; verify that table-crash command proposals are actively vetoed by the STM32.
• Weekly Deliverable: UnoQMotorsBus.py source code + demonstration of live table-crash veto.

---

Sprint 2: Sensing, Data Collection & Policy Baseline (Weeks 5–7)

Focus: Sensor calibration, human demonstration recording, and baseline imitation policy training.

Week 5 (Oct 19 – Oct 23, 2026): “Student Zero” Run for Labs 1 & 2

• Primary Objective: Qualify Lab 1 (Causal Boundary) and Lab 2 (Sensing & Bridge).
• Tasks:
  • Mount UVC USB webcam at \(45^\circ\) oblique angle, \(40\text{ cm}\) overlooking workspace; tape \(300 \times 200\text{ mm}\) boundary.
  • Write and test AprilTag camera calibration script (calibrate_camera.py).
  • Write joint homing calibration script; verify \(\pm 1.5^\circ\) repeatability across 20 cycles.
  • Benchmark inter-core RPC latency: measure round-trip times across 1,000 packets; confirm \(\le 5\text{ ms}\) at \(50\text{ Hz}\).
  • Package student starter assets: wiring_diagram_golden.pdf, calibrate_camera.py, test_bridge_latency.py.
• Weekly Deliverable: Milestone 1 qualification package (Lab 1 & 2 starter repo + reference calibration card).

Week 6 (Oct 26 – Oct 30, 2026): “Student Zero” Run for Lab 3 (Teleoperation & Datasets)

• Primary Objective: Record golden demonstration dataset in standardized LeRobot v2 schema.
• Tasks:
  • Establish reproducible pick-and-place task: pick \(30\text{ mm}\) foam cube from marked start zone and place in receptacle.
  • Record 10 expert demonstration episodes using teleoperation.
  • Synchronize \(30\text{ Hz}\) RGB frames (\(224 \times 224\)) with the 4 action taps (a_req, a_map, a_enf, a_meas).
  • Format into LeRobotDataset v2 schema; verify zero dropped frames and correct chunk indexing.
  • Package student starter scripts: teleop_record.py, inspect_dataset.py, dataset_golden_10ep/.
• Weekly Deliverable: Published 10-episode reference dataset with complete dataset card and inspection plots.

Week 7 (Nov 2 – Nov 6, 2026): “Student Zero” Run for Lab 4 & Milestone 2

• Primary Objective: Train baseline ACT policy, quantize to ONNX INT8, and build heuristic baseline.
• Tasks:
  • Implement deterministic heuristic reach script (heuristic_reach.py) for baseline comparison.
  • Train Action Chunking with Transformers (ACT) model on workstation/Colab using the 10 golden episodes (\(\le 45\text{ min}\) training).
  • Quantize trained PyTorch model to ONNX INT8 format (export_onnx_quantized.py).
  • Verify model file size is \(< 50\text{ MB}\) and validates on held-out evaluation episodes.
  • Package starter assets: train_act_baseline.py, pretrained_act_int8.onnx, heuristic_reach.py.
• Weekly Deliverable: Milestone 2 qualification package (trained checkpoint + offline MSE evaluation curves).

---

Sprint 3: Untethered Edge Deployment & Dynamic Adaptation (Weeks 8–10)

Focus: Natively running neural policies on Qualcomm Linux, action chunk horizons, and closed-loop disturbance recovery.

Week 8 (Nov 9 – Nov 13, 2026): Validation Step 4 & Lab 5 (Untethered Autonomous Reach)

• Primary Objective: Run closed-loop visual reach completely untethered on the Arduino UNO Q.
• Tasks:
  • Transfer policy_int8.onnx and pai_edge_runtime.py to Qualcomm Linux storage.
  • Unplug host PC tether; board runs exclusively on 45W USB-PD supply with webcam and governed arm.
  • Launch autonomous runtime via SSH over Wi-Fi: \[\text{Webcam Frames } (30\text{ Hz}) \longrightarrow \text{ONNX Policy } (<45\text{ ms}) \longrightarrow \text{RPC Bridge } (<5\text{ ms}) \longrightarrow \text{STM32 Motion}\]
  • Validation Step 4 Check: Verify end-to-end loop latency \(T_{\text{total}} \le 80\text{ ms}\); arm autonomously reaches foam block at arbitrary locations across 10 trials.
  • Package starter assets: pai_edge_runtime.py, telemetry_logger.py, sample_reach_trace.csv.
• Weekly Deliverable: Step 4 verification report + multi-tap telemetry CSV of untethered reach.

Week 9 (Nov 16 – Nov 20, 2026): “Student Zero” Run for Lab 6 (Action Horizons & Disturbances)

• Primary Objective: Evaluate action chunk horizons and closed-loop disturbance recovery.
• Tasks:
  • Benchmark action chunk horizons: \(K \in \{1, 8, 16, 32\}\). Measure tracking error and physical trajectory smoothness.
  • Test mid-trajectory disturbance: shift target block by \(50\text{ mm}\) during open-loop chunk execution; measure recovery time.
  • Evaluate language conditioning: run test prompts (“pick red block” vs “pick blue block”) on SmolVLA checkpoint; verify trajectory divergence.
  • Package starter assets: benchmark_chunking.py, disturbance_eval.py, smolvla_eval_harness.py.
• Weekly Deliverable: Milestone 3 qualification package (horizon comparison plots + disturbance recovery video).

Week 10 (Nov 23 – Nov 27, 2026): Mid-Term Rig Reliability & “Student Zero” Run for Lab 7 (MCU Safety Governor)

• Primary Objective: Conduct continuous stress testing and program the real-time STM32 safety governor.
• Tasks:
  • Run 50 continuous autonomous pick-and-place cycles on the Golden Station.
  • Measure STS3215 servo casing temperatures on shoulder (\(J_2\)) and elbow (\(J_3\)) joints; verify temperature stays below \(60^\circ\text{C}\).
  • Profile V4L2 camera buffer over 2 hours of continuous streaming; confirm zero memory leaks or dropped frame buffers.
  • Finalize STM32 firmware for joint velocity saturation (\(\omega_i \le 45^\circ/\text{s}\)) and Cartesian floor geofence (\(z \ge 15\text{ mm}\)).
  • Write Python fault injection script (inject_table_crash.py) that sends deliberate downward velocity spikes and out-of-bound targets.
  • Demonstrate real-time MCU veto: verify that illegal proposals are rejected in \(< 5\text{ ms}\) while legal joints continue moving safely.
  • Package starter assets: stm32_governor_firmware/, inject_table_crash.py, veto_audit_golden.csv.
• Weekly Deliverable: 50-cycle reliability & thermal profile report + Lab 7 qualification package (firmware source + live veto audit).

---

Sprint 4: Real-Time Governance & Fault Hardening (Week 11)

Focus: Microcontroller safety boundary enforcement, hardware watchdogs, and zero-backlog recovery.

Week 11 (Nov 30 – Dec 4, 2026): “Student Zero” Run for Lab 8 & Milestone 4

• Primary Objective: Hardware watchdog timeout and fail-safe cutoff under system crash.
• Tasks:
  • Program \(150\text{ ms}\) hardware SysTick watchdog timer in STM32 firmware.
  • Inject Linux process freeze (kill -STOP <pid>) during active arm motion.
  • Verify that the arm decelerates to a safe stop within \(150\text{ ms}\) of heartbeat loss.
  • Resume Linux process (kill -CONT <pid>); verify that the MCU flushes stale UART queues and executes zero stale commands.
  • Package starter assets: stm32_watchdog_firmware/, fault_injection_suite.py, recovery_verification.py.
• Weekly Deliverable: Milestone 4 complete qualification package (watchdog timing trace + zero-backlog proof).

---

Sprint 5: Capstone Rehearsal & Class-Set Duplication (Weeks 12–13)

Focus: Capstone defense simulation, golden disk image freeze, and independent bench replication.

Week 12 (Dec 7 – Dec 11, 2026): Capstone Rehearsal & 20-Trial Release Protocol

• Primary Objective: Simulate the final Capstone Milestone 5 defense protocol.
• Tasks:
  • Define official Capstone evaluation benchmark: 20 witnessed physical trials across held-out target locations and surface obstacles.
  • Execute 20 consecutive trials as “Student Zero”; record success rate (target \(\ge 80\%\)) and safety violations (must be \(0\%\)).
  • Prepare Claim-Argument-Evidence (CAE) Physical Release Dossier template for students.
  • Audit complete student-facing rubric against the 12 competencies (A1–D3).
• Weekly Deliverable: Reference Capstone Physical Release Dossier + 20-trial unedited video archive.

Week 13 (Dec 14 – Dec 18, 2026): Golden Image Freeze & Independent Replication Audit

• Primary Objective: Freeze software images and have a second staff member replicate Station 2 from scratch.
• Tasks:
  • Build Golden Qualcomm Linux SD card image: pre-installed Debian, LeRobot v0.2+, PyTorch ARM64, ONNX Runtime, V4L2 utilities, pre-cloned course repository.
  • Freeze STM32 pre-flashed binary for one-click flashing via Arduino App Lab.
  • Independent Replication Test: Have a secondary researcher or teaching assistant unbox a second UNO Q and SO-101 arm, follow the staff setup guide, flash the images, and achieve standalone reach.
  • Assemble spare parts buffer: 4x spare STS3215 servos, 2x spare webcams, replacement 3D-printed brackets, multimeters.
  • Final sign-off review: certify lab readiness for Spring 2027 studio launch.
• Weekly Deliverable: Golden .img file uploaded to lab server + signed Replication Audit Report.

---

5. Master Deliverables & Weekly Technical Sign-Off Checklist

Week & Date Range	Milestone / Sprint Phase	Primary Physical AI Deliverables	Hardware Verification & Acceptance Criteria	Status
W01: Sep 21 – Sep 25	Master Plan & Prep	• Complete curriculum review & syllabus freeze • BOM inventory & receiving audit • Dedicated bench allocation & power rail check	Multimeter confirmation of 7.4V DC & 45W USB-PD rails; complete toolchain staged	[ ]
W02: Sep 28 – Oct 02	Hardware Bring-Up	• Mechanical arm assembled & clamped • 7.4V motor power harness verified • Step 1: LeRobot USB teleoperation passing	60-second teleoperation replay executed without communication dropout	[ ]
W03: Oct 05 – Oct 09	Bus Authority	• STM32 UART level-shifter circuit wired • Inter-core RPC bridge test script • Step 2: 1-joint velocity clamp & veto	Joint 1 clamped at \(45^\circ/\text{s}\); out-of-range targets (\(> 180^\circ\)) rejected	[ ]
W04: Oct 12 – Oct 16	6-DoF Integration	• UnoQMotorsBus Python adapter deployed • Table geofence (\(z \ge 15\text{ mm}\)) programmed • Step 3: Governed 6-DoF teleoperation	Full 6-DoF teleoperation functional; downward table crash actively vetoed	[ ]
W05: Oct 19 – Oct 23	Milestone 1: Anatomy	• Lab 1 (Boundary) qualified by Student Zero • Lab 2 (Sensing & Bridge) qualified • AprilTag camera & joint offset calibration	Joint repeatability \(\pm 1.5^\circ\); inter-core bridge round-trip \(\le 5\text{ ms}\) at \(50\text{ Hz}\)	[ ]
W06: Oct 26 – Oct 30	Dataset Pipeline	• Lab 3 (Teleoperation) qualified • 10-episode pick-and-place reference dataset • Standardized LeRobot v2 schema validated	Zero dropped frames at \(30\text{ Hz}\); all 4 action taps synchronized	[ ]
W07: Nov 02 – Nov 06	Milestone 2: Training	• Lab 4 (Baseline & Export) qualified • ACT model trained on workstation GPU • Quantized ONNX INT8 policy exported (\(< 50\text{ MB}\))	Offline MSE validation curves clean; model file size \(< 50\text{ MB}\)	[ ]
W08: Nov 09 – Nov 13	Edge Inference	• Step 4: Untethered reach on Qualcomm Linux • End-to-end loop latency \(T_{\text{total}} \le 80\text{ ms}\) • Lab 5 (Autonomous Reach) qualified	Host PC tether disconnected; 10/10 autonomous reach completions	[ ]
W09: Nov 16 – Nov 20	Milestone 3: Dynamics	• Lab 6 (Action Horizons) qualified • Horizon benchmark curves (\(K=1..32\)) • Disturbance recovery mid-trajectory verified	Measurable trajectory adaptation under \(50\text{ mm}\) block displacement	[ ]
W10: Nov 23 – Nov 27	Stress & Governor	• 50-cycle continuous autonomous reliability test • Servo thermal profiling (\(< 60^\circ\text{C}\)) • Lab 7 (Safety Governor) qualified	Servo casing \(< 60^\circ\text{C}\); live table-crash proposals vetoed in \(< 5\text{ ms}\)	[ ]
W11: Nov 30 – Dec 04	Milestone 4: Cutoffs	• Lab 8 (Watchdog Cutoffs) qualified • \(150\text{ ms}\) hardware SysTick watchdog verified • Linux process freeze & zero backlog proved	Arm halts in \(\le 150\text{ ms}\) upon heartbeat loss; zero stale commands on resume	[ ]
W12: Dec 07 – Dec 11	Capstone Rehearsal	• Capstone Studio protocol simulated • 20 witnessed physical trials executed • Claim-Argument-Evidence dossier template	\(\ge 80\%\) task success rate across 20 trials with zero safety violations	[ ]
W13: Dec 14 – Dec 18	Final Release Freeze	• Golden Linux SD card image compiled • STM32 baseline binary locked • Secondary TA replication test completed	Independent replication on Station 2 passes standalone reach without staff intervention	[ ]

---

6. Weekly Execution & Reporting Protocol

To ensure rapid resolution of hardware bugs and maintain transparent progress toward the December deadline:

1. Weekly Friday Check-In (Every Friday by 17:00 CET):
   • Submit a structured weekly update via GitHub Issue / PR containing:
     • Status: [ON TRACK] / [DELAYED] / [BLOCKED]
     • Completed Tasks: Checkboxes matching the weekly sprint table above.
     • Evidence Links: Commit hash, telemetry CSV, model evaluation curve, or uncut video clip.
     • Active Blockers: Detailed diagnostic logs of any hardware/timing failure.
2. Bi-Weekly Physical Sync:
   • 30-minute bench review demonstrating the hardware bring-up steps and milestone deliverables on live hardware.
3. Escalation Trigger:
   • If any bring-up step (Steps 1–4) is delayed by more than 4 business days, activate the pre-approved fallback matrix in feasibility-plan.md immediately.