Repair Intelligence
The bridge between human technicians, robot hardware, and AI systems. Study failure modes, understand component anatomy, explore simulation environments, and take your certification exam — all from one place.
Platforms
20
Failure Signatures
73
Sim Environments
6
AI Layers
4
Layer 1 — Physical
Each platform entry includes known failure signatures, severity, and maintenance context. Expand any platform to study its repair profile before your exam.
Humanoid
Unitree Robotics
43-DOF bimanual humanoid with Dex1 hands. VLA-capable via UnifoLM-VLA-0. Official TechMedix partner platform.
Total DOF
43
Arm DOF
7 per arm
F/T Sensors
13
Control Freq
20 Hz
Known Failure Signatures
Actuator Overheat
Joint temp > 75°C sustained > 30s
Joint Backlash
EEF tracking error > 15mm
F/T Sensor Drift
Static FT reading > ±2N / ±0.5Nm
Unitree Robotics
Full-size humanoid at 180cm. 40-DOF, extended battery for long-shift warehouse and logistics deployment.
Total DOF
40
Height
180 cm
Payload
30 kg
Battery
~4h operation
Known Failure Signatures
Actuator Overheat
Leg actuator temp > 80°C during sustained gait
Joint Backlash
Leg joint position error > 20mm
Battery Critical
SOC < 15% during active task
Figure AI
Commercial humanoid for automotive assembly. 16 DOF hands, OpenAI-trained reasoning. BMW Group deployment.
Height
168 cm
Weight
70 kg
Battery
~5h
Hand DOF
16 per hand
Known Failure Signatures
Hand Gear Wear
Finger position error > 2mm on precision grip
Head Camera Drop
Primary vision system offline > 100ms
Tesla
Tesla's 3rd-gen humanoid. 22 DOF hands, FSD-derived perception stack, designed for Tesla factory operations.
DOF
35+
Weight
57 kg
Battery
~8h
Hand DOF
22 per hand
Known Failure Signatures
FSD Stack Latency
Perception inference > 50ms
Battery Critical
SOC < 20% during active sequence
Agility Robotics
Warehouse-specialized bipedal. Deployed at Amazon. Optimized for tote handling in dynamic logistics environments.
Payload
16 kg
Height
175 cm
Battery
~16h (hot-swap)
Speed
1.5 m/s
Known Failure Signatures
Ankle Backlash
Ankle dorsiflexion error > 1.5°
Knee Actuator Overheat
Knee joint > 76°C during incline traversal
Physical Intelligence
π0-powered general-purpose humanoid. Diffusion policy control. Research-to-production platform.
Policy
π0 Diffusion
DOF
38
Battery
~3h
Known Failure Signatures
Diffusion Stall
Policy inference > 100ms per step
Gripper Drift
End-effector grip force variance > 15N
Industrial
Unitree Robotics
Industrial quadruped. 23 kg payload, IP67, outdoor-capable. Common in construction survey and infrastructure inspection.
Payload
23 kg
Protection
IP67
Battery
~4h
Speed
6 m/s
Known Failure Signatures
Leg Actuator Overheat
Leg joint > 72°C on rough terrain
Foot F/T Drift
Contact detection false positive rate > 5%
Boston Dynamics
Industry-standard inspection quadruped. Arm-optional, API-first, payload-extensible. Deployed in 50+ industries.
Payload
14 kg
Battery
~90 min
Protection
IP54
Speed
1.6 m/s
Known Failure Signatures
Hip Actuator Wear
Hip position error > 0.8° during gait
Foot Sensor Drift
Ground contact confidence < 85%
Amazon Robotics
Autonomous mobile robot for warehouse pod movement. LiDAR-only navigation, no camera dependence for safety.
Payload
750 kg (pod)
Speed
1.1 m/s
Battery
~8h (hot-swap)
Nav
LiDAR SLAM
Known Failure Signatures
Drive Motor Overheat
Drive motor > 70°C under heavy pod
LiDAR Dropout
LiDAR scan rate < 5 Hz
Seeed Studio
True open-source 6-DOF robot arm — hardware blueprints, full BOM, and 3D print files published on GitHub. Python SDK, ROS1, ROS2, Isaac Sim, and LeRobot compatible. Same simulation stack as ROBO-1. Pre-integration: monitoring adoption curve for full TechMedix onboarding.
DOF
6
SDK
Python, ROS1, ROS2, Isaac Sim, LeRobot
Source
Open — hardware blueprints, full BOM, 3D print files
GitHub
Seeed-Projects/reBot-DevArm
Known Failure Signatures
Joint Backlash
EEF position error > 2mm on repetitive trajectory
Servo Overheat
Joint servo > 70°C during sustained high-torque moves
2 signatures
NVIDIA
Flagship edge AI compute module for humanoid robots, surgical systems, and industrial AI. Blackwell GPU, 128GB LPDDR5X, 40-130W. Ships with JetPack 7 / Ubuntu 24.04 / CUDA 13.0. Developer kit $3,499, GA August 2025.
AI Compute
2,070 TFLOPS FP4
Memory
128GB LPDDR5X
GPU
Blackwell
Power
40-130W
Known Failure Signatures
JetPack/CUDA Driver Conflict
CUDA 13.0 driver mismatch after Orin-to-Thor migration breaks inference stack -- verify JetPack 7 compatibility before migration
NIM Microservice Cold Start
Cosmos Reason NIM container latency >5s on first inference call at edge -- pre-warm containers on boot
Thermal Throttle
Sustained load >100W triggers clock throttling -- verify thermal paste application and enclosure airflow
6 signatures
Aigen
100% solar-powered agricultural robot for chemical-free weed control at plant level. 350W panel, all-wheel drive, stereo depth vision (gen2), intelligent mesh fleet coordination. Featured by NVIDIA Robotics (April 2026). Deployed in California Central Valley.
Power
Solar + battery (350W)
Drive
All-wheel drive
Field Hrs
Up to 14h/day
Fleet Cov
200 acres/season
Known Failure Signatures
Weed Misclassification
Dense cotton canopy (late season) causes false negative rate >15% -- retrain vision model with seasonal field data
Mesh Network Dropout
Fleet communication loss in hilly terrain -- robot becomes unreachable, manual retrieval required
Solar Charging Shortfall
Multi-day overcast reduces daily operational range by 40-60% -- pre-charge backup batteries before extended cloudy periods
4 signatures
Drone
DJI
50L agricultural spray drone. Dual atomized + broadcast spreading. Active phased-array radar obstacle avoidance.
Payload
50 L / 40 kg
Spread Width
9 m
Flight Time
~17 min (full load)
Radar
Active phased-array
Known Failure Signatures
Motor Overheat
Prop motor > 85°C (full payload)
Battery Critical
SOC < 15% mid-mission
Skydio
Enterprise inspection drone with NVIDIA Jetson-based autonomy and 4K/thermal dual sensor. Used in infrastructure inspection.
Camera
4K + thermal
Flight Time
35 min
Wind Resistance
15 m/s
Autonomy
NVIDIA Jetson
Known Failure Signatures
Thermal Sensor Drop
Thermal feed missing > 500ms
Autonomy Stack Stall
Obstacle avoidance latency > 200ms
Zipline
Fixed-wing VTOL delivery drone. 160km range, autonomous dock-to-dock. Medical and commercial delivery.
Range
160 km
Payload
3.6 kg
Speed
128 km/h
Type
Fixed-wing VTOL
Known Failure Signatures
Tilt Motor Overheat
VTOL tilt motor > 90°C during transition
Range Battery Low
SOC < 25% beyond return threshold
Delivery
Serve Robotics
Sidewalk delivery robot. LiDAR + camera fusion, 4-wheel drive. Deployed in LA and Dallas for food delivery.
Speed
1.5 m/s
Payload
18 kg
Battery
~12h
Sensors
LiDAR + 8× camera
Known Failure Signatures
Perimeter Camera Drop
Any of 8 cameras offline > 200ms
Wheel Encoder Drift
Odometry error > 5cm/m traveled
Starship Technologies
Campus delivery robot. 6-wheel drive, 40+ cm obstacle clearance. Operates in 50+ cities.
Speed
1.8 m/s
Payload
10 kg
Battery
~6h
Cameras
12× wide-angle
Known Failure Signatures
Battery Critical
SOC < 12% away from dock
Drive Motor Overheat
Wheel motor > 65°C on incline
Micromobility
Lime
4th-gen shared e-bike. Swappable battery, IoT-connected, deployed in 200+ cities.
Motor
250W rear hub
Battery
36V swappable
Range
40+ miles
Connectivity
Proprietary cellular
Known Failure Signatures
Battery Cell Drift
±50mV cell delta — pack imbalance, swap recommended
Battery Thermal
>45°C at rest — thermal event risk
Hub Motor Bearing
Vibration >0.8g RMS — bearing wear onset
Bird
3rd-gen shared scooter. Reinforced frame, 40km range, GPS geofencing. 350W motor with integrated non-swappable battery.
Motor
350W rear hub
Battery
36V 10.5Ah integrated
Range
25+ miles
Connectivity
Cellular + BLE
Known Failure Signatures
Battery Cell Drift
±50mV cell delta — pack imbalance
Battery Thermal
>60°C BMS cutoff — non-swappable pack at risk
Hub Motor Bearing
Vibration >0.8g RMS — bearing wear onset
Rad Power Bikes
Commercial cargo e-bike. 150kg total load capacity, 750W motor, used for last-mile delivery fleets.
Motor
750W rear hub
Battery
48V 14Ah 672Wh
Range
45+ miles
Payload
120 lbs
Known Failure Signatures
Battery Cell Drift
±60mV cell delta — large 48V pack imbalance
Battery Thermal
>50°C at rest — 672Wh pack thermal event risk
Hub Motor Bearing
Vibration >0.8g RMS under cargo load
Layer 1 — Physical
What is physically inside a robot — and how each part fails. Every entry includes a "human bridge" — how to feel, see, or measure the failure before you ever open a diagnostic tool.
A typical humanoid uses 25-30 actuators split between rotary (shoulders, elbows, hips, knees) and linear (legs, torso). Each rotary actuator combines BLDC motor + harmonic reducer + encoder + torque sensor.
Key Fact
Tesla Optimus: 20 rotary + 14 linear actuators. Cost breakdown: reducer 36%, torque sensor 30%, motor 13.5%.
Human Bridge
Think of actuators as the robot's muscles -- they create movement at every joint. When an actuator fails, the robot loses range of motion in that limb.
Strain wave reducers provide high gear ratios with near-zero backlash. Largest single cost driver (~36%) in rotary actuators. Only 12% of global machine tool manufacturers meet the precision requirements.
Key Fact
Harmonic Drive holds 20-25% global market share. Alternatives: cycloidal-pin gear, planetary gearbox.
Human Bridge
Reducers are why robots move smoothly rather than jerking. Backlash (slop in the gears) is a primary failure symptom -- you'll feel it as imprecision or vibration at the end-effector.
Brushless DC motors are the dominant choice across humanoids for high torque density in compact form factors. Controlled by ESC/FOC drivers. No brushes = longer service life vs brushed motors.
Key Fact
Dominant suppliers: Maxon (25-30% share), Kollmorgen (15-20%). China alternative: PMSM low-inertia high-speed motors (Unitree).
Human Bridge
If a joint runs hot or draws unusual current, the motor winding or controller (ESC) is your first diagnostic target. Oscilloscope the drive signal before replacing the motor.
Largely standardized on NVIDIA Jetson (Orin, AGX Thor). Tesla is the exception with its proprietary AI5 SoC. Key metrics: TOPS, TOPS/watt, and memory bandwidth.
Key Fact
NVIDIA Jetson AGX Thor: 2,070 FP4 TFLOPS (Blackwell). Tesla AI5: proprietary. Chinese alternative: Horizon Robotics.
Human Bridge
The compute module is the robot's brain. Heat + throttling = inference slowdown = sluggish reactions. Check thermal paste and airflow before suspecting software issues.
2-7 cameras per robot for perception. IMUs for orientation. Force/torque sensors on joints. LiDAR for mapping (Agility, Unitree, AGIBot). Tactile sensors on hands for dexterous manipulation.
Key Fact
10 of 13 major OEMs have tactile sensing. Sony and Intel RealSense dominate camera supply.
Human Bridge
A camera dropout causes immediate autonomy loss. IMU drift causes balance instability. F/T sensor miscalibration causes grip failures. Each sensor has a clear, testable failure signature.
Lithium-ion and lithium-polymer packs. Operating times range 2-14 hrs. Cell imbalance (+/-50mV delta) is the primary aging indicator. XPeng is pushing toward all solid-state.
Key Fact
Capacity range: 0.84-5 kWh across humanoids. Charging modes: self-charge, hot-swap, wireless inductive.
Human Bridge
Battery health is the most measurable robot health signal. Monitor cell delta voltage, not just total SOC. A swollen pack is a thermal event risk -- never charge it.
Range from 3-finger grippers to 22-DOF anthropomorphic hands. Drive types: tendon (1X, Tesla) vs motor+gear (Figure, Dexmate). Research standard: Shadow Robot 24 DOF, ~110K EUR.
Key Fact
Open-source option: ORCA Dexterity (17 DOF, $3.5K-$6.1K) from ETH Zurich spinoff.
Human Bridge
Hand repairs are the most tactile diagnostic task. Encoder drift, tendon tension loss, and fingertip sensor fouling each have distinct feel and response pattern. L2+ certification required.
ISO 25785-1 (bipedal robots) is in working draft -- expected 2026-2027. ISO 10218 covers industrial robots. ISO 13482 covers service robots. EU Machinery Regulation applies Jan 2027.
Key Fact
Only Agility Robotics Digit has achieved NRTL field certification. AI Act applies to high-risk autonomous systems from August 2026.
Human Bridge
LOTO (Lock Out Tag Out) and zero-energy verification before any physical work. Know which standard covers the robot you're servicing -- it determines your liability and documentation requirements.
Layer 2 — Intelligence
Where humans get familiar with robot parts, failure scenarios, and toolsets before working on real hardware. These environments let you practice repairs, run fault injection, and understand how AI policies respond to component failures.
Train humanoid locomotion and manipulation policies. Run thousands of parallel simulations. Integrates with ROS 2 and real hardware via Isaac ROS.
Engine
PhysX + Newton + MuJoCo
License
Free for R&D
Language
Python, C++
Cert Level
L3–L5
Fast contact physics. Standard benchmark environment for locomotion RL. Google DeepMind maintains it. MJX runs on GPU via JAX for massively parallel training.
Engine
MuJoCo (JAX backend via MJX)
License
Apache 2.0
Language
C, Python, JAX
Cert Level
L2–L4
26 seconds to train a locomotion policy on an RTX 4090. Supports rigid, soft, fluid, and cloth simulation. Best for rapid policy iteration and diverse material interaction.
Engine
Custom multi-physics (rigid, MPM, SPH, FEM, PBD)
License
Apache 2.0
Language
Python
Cert Level
L3–L5
Benchmark + training framework where LLM/VLM agents write Python control code to operate physical robots. 187 tasks across Robosuite, LIBERO-PRO, and BEHAVIOR. CaP-RL takes a 7B model from 20% to 72% success in 50 iterations. 84% sim-to-real transfer on Franka Panda. Maps to TechMedix cert levels: L1=S1 tasks, L2=S2-S3+M1-M2, L3=S4+M3-M4.
Engine
MuJoCo + Isaac Sim (BEHAVIOR tasks)
License
Open Source
Language
Python
Cert Level
L1–L3
Under evaluation for TechMedix fault injection simulation track. Target use: technicians diagnose simulated hardware failures in a virtual environment before working on real robots.
Engine
Python / ROS 2 (evaluation)
License
MIT (evaluation)
Language
Python
Cert Level
L1–L2
NVIDIA + Google DeepMind + Disney Research convergence project. Bridges Isaac Sim and MuJoCo ecosystems. Expected to become the shared physics layer for robot training.
Engine
Newton (on NVIDIA Warp)
License
Open Source (Linux Foundation)
Language
Python, C++
Cert Level
L4–L5
Layer 2 — Intelligence
How robots think, learn, and generalize — and what it means for diagnostics. Understanding the AI stack helps technicians know whether a failure is hardware, firmware, or policy-level.
Learn to predict how the physical world evolves in response to robot actions. Let robots practice tasks "in imagination" before acting in reality — dramatically reducing real-world data needs.
Examples
NVIDIA Cosmos, Wayve GAIA, Dreamer v3
Tech-to-Field Bridge
When a robot hesitates before a new task, it may be running a world model rollout to predict outcomes. Failure signatures: policy stalls, repetitive motions, unexpected stops.
Vision-Language-Action models unify perception, language grounding, and control policies. They are the generalist reasoning layer — the robot's 'common sense' for embodied tasks.
Examples
π0 (Physical Intelligence), OpenVLA, RoboFlamingo, Helix (Figure), UnifoLM-VLA (Unitree)
Tech-to-Field Bridge
VLA inference stall = motor overheat or perception dropout feeding corrupted state. Check camera feeds and motor temps before blaming the model.
Score video trajectories against language instructions. Provide the dense reward signal that VLA policies need for RL training, data filtering, and quality estimation.
Examples
Value-Order Correlation (VOC) scorers, trajectory preference models
Tech-to-Field Bridge
During policy training, reward model quality directly determines task success rate. A miscalibrated reward model produces robots that "game" the score without completing the task.
The sim-to-real gap is the central challenge. Domain randomization (varying friction, lighting, mass) in simulation trains policies that transfer to real hardware.
Examples
Isaac Lab DR, MuJoCo domain randomization, Genesis material variation
Tech-to-Field Bridge
Policies that work in sim but fail on hardware usually have a sensor calibration mismatch or a joint friction model error. The fix is hardware measurement, not re-training.
Layer 3 — Human
Five certification levels from Operator to Autonomous Systems Architect. Study the guide above, then take the exam directly — no GitHub required.