06Reference framework
A common language for robotics beyond the factory
Four axes to classify any non-industrial robotic system (humanoids, quadrupeds, autonomous vehicles and drones) by technology maturity, operational safety and return on investment.
Last reviewed: 2026-11Methodology & sources →
01 · Scope & overview
Robotics for places nobody designed for robots
The framework covers platforms that operate outside controlled industrial settings: extreme dynamics, unstructured infrastructure and hostile or changing conditions. We classify progress across four physical forms.
- HUM
Humanoids
Bipedal systems with dexterous manipulation, built to use tools and infrastructure designed for humans.
4 in the explorer
- QDR
Quadrupeds
Articulated legged platforms with high traction and stability on uneven terrain, stairs and rock.
5 in the explorer
- UGV
Ground vehicles
Wheeled or tracked systems for cargo transport, logistics and long-range inspection.
4 in the explorer
- UAV
Drones & aerial robots
Rotary or fixed wing for monitoring, vertical inspection, express logistics and swarms.
5 in the explorer
02 · Strategic objectives
From the vendor catalog to the operational need
Close the operational gap
Connect what robotics vendors can do with what operations actually need.
Standardize the criteria
A shared vocabulary to assess any platform with three questions:
- Maturity
- Which autonomy level does it sustain, and in which operational domain?
- Safety
- When and how does a human step in, and what happens if the link is lost?
- ROI
- When does it pay back against manual operation, including total cost of ownership?
The four axes
Every robot is described by four coordinates
They are categorical and independent dimensions: a drone can have high autonomy and no coordination, and a humanoid can have rich perception and teleoperated control.
Axis 1 · Autonomy & decision making
Six levels, from teleoperation to distributed autonomy
The level describes the typical operational autonomy in real deployments, not the brochure maximum. Select a level to see which platforms in the explorer sit there today.
This 0–5 scale is a framework adapted by this site for non-industrial robotics. It draws on the SAE J3016 levels of driving automation and the ODD concept, but it is not an official standard and does not replace certification.[ANSI Blog, 2021][ISO, 2021]
L4
High autonomy
Operates continuously without human intervention inside a bounded operational design domain (ODD).
Platforms at this level
Key: the operational design domain (ODD)
Level 4 only makes sense within an explicit ODD: geography, weather, lighting, obstacle types and speeds. The same robot can be L4 in a known substation and L2 on new terrain.[Fraunhofer IKS]
Axis 2 · Networking & multi-robot coordination
From the isolated robot to swarm intelligence
The communication topology determines scale, resilience and what happens when a node fails.
Standalone
5 platformsNo inter-robot communication; each unit operates in isolation.
- No network dependency
- Scale = number of operators
- Isolated failure
Coordinated
13 platformsShares state and telemetry through a base station or central server.
- Base station or cloud
- Central fleet planning
- Single point of failure
Collaborative swarm
0 platformsDecentralized decisions over a peer-to-peer mesh; tasks are allocated by each node's capability and position.
- Peer-to-peer mesh / MANET
- Dynamic task allocation
- Graceful degradation
Axis 3 · Perception & environment observation
No sensor sees everything: fusion is the key
Every modality has a blind spot. Mature platforms combine several to keep localizing and detecting in darkness, fog, dust or without GNSS.
3D LiDAR
10/18- + Adds:
- Precise 3D geometry, day and night
- − Limit:
- Cost; degraded by dense fog and rain
Stereo cameras / VIO
6/18- + Adds:
- Lightweight, the basis of GPS-free VIO
- − Limit:
- Needs light and texture
RGB-D depth cameras
5/18- + Adds:
- Dense depth at short range
- − Limit:
- Short range, sunlight-sensitive
RGB cameras
11/18- + Adds:
- Rich semantics for vision models
- − Limit:
- No direct depth; fails in darkness
Thermal / infrared
7/18- + Adds:
- Sees heat: people, leaks, hot spots
- − Limit:
- Low resolution, no color or texture
mmWave radar
2/18- + Adds:
- Sees through fog, dust and smoke; measures velocity
- − Limit:
- Low angular resolution
GNSS / RTK
9/18- + Adds:
- Centimeter global position with RTK
- − Limit:
- Useless underground or under jamming
IMU
18/18- + Adds:
- High-rate motion, always available
- − Limit:
- Drift grows without corrections
2D LiDAR
0/18- + Adds:
- Cheap and reliable for indoor planes
- − Limit:
- Single plane: misses high or low obstacles
Tactile / haptic sensors
1/18- + Adds:
- Force and slip during manipulation
- − Limit:
- Only informs on contact
Ultrasonic
1/18- + Adds:
- Cheap proximity, transparent surfaces
- − Limit:
- Short range, low precision
Acoustic / leak ultrasound
2/18- + Adds:
- Gas or air leaks and bearing faults
- − Limit:
- Industrial ambient noise
Gas detection
0/18- + Adds:
- Hazardous gas concentration
- − Limit:
- Point measurement, needs calibration
PTZ zoom camera
2/18- + Adds:
- Optical zoom to read gauges from afar
- − Limit:
- Narrow field of view
Sensor fusion & SLAM
Simultaneous localization and mapping (SLAM) lets a robot know where it is without GPS. The typical pipeline has five stages:
- 01Sensors
LiDAR, cameras, IMU, wheel or leg odometry
- 02Front-end
Feature extraction and visual-inertial (VIO) or LiDAR odometry
- 03Filter
Extended or unscented Kalman filter (EKF/UKF) fusing estimates at high rate
- 04Back-end
Factor-graph optimization and loop closure to correct drift
- 05Map + pose
A consistent 3D map and the robot's position in it
GPS-denied localization
In tunnels, mines, indoors or under interference, GNSS disappears. VIO estimates motion by comparing successive frames with the IMU; LiDAR SLAM aligns point clouds; and loop closure recognizes previously visited places to cancel accumulated drift.
Axis 4 · Control & operation modes
From the FPV pilot to the learned policy
The control mode tells how much of the behavior is written by an engineer, executed by a pilot or learned by a neural network.
- 01
FPV piloting
Low-latency video links (<50 ms) for maneuvers in tight spaces. The pilot is the perception system.
1/18 - 02
Teleoperation
The robot stabilizes and protects itself; the human decides every move. The basis of data collection for imitation learning.
15/18 - 03
Supervised mission
Recorded or scheduled missions the robot repeats; the operator approves exceptions from a console.
14/18 - 04
BVLOS fleet
One operator supervises several units beyond visual line of sight. Requires detect-and-avoid and regulatory authorization.
4/18 - 05
Learned policy (RL/VLA)
Networks trained with reinforcement learning or vision-language-action models, often with physics-informed neural networks (PINN) for gait and manipulation.
8/18
Sim-to-real: train in simulation, deploy in the world
Locomotion policies are trained for millions of steps in parallel simulators with domain randomization (friction, masses, latencies, sensor noise) so they survive reality. The residual gap is closed with field fine-tuning and teleoperation data.
- Massive simulation
- Domain randomization
- Deployment
- Field tuning
Putting it to work
How to apply the taxonomy when assessing a vendor
Each axis translates into verifiable questions. Ask for evidence of real operation in a domain similar to yours, not staged demos.
- Autonomy
Question: In which specific operational design domain (ODD) has the claimed level been demonstrated?
Evidence to request: Real mission logs with human interventions per hour and their causes, under conditions comparable to yours.
- Coordination
Question: What does each robot do if it loses the link to the base station or the cloud?
Evidence to request: Documented disconnection behavior: which tasks continue, which are suspended and how the mission resumes.
- Perception
Question: Which sensors are in the configuration being offered, not the catalog, and how does it localize without GNSS?
Evidence to request: Tests with dust, rain, low light or reflections and the localization accuracy achieved at your site.
- Control
Question: Which part of the mission is teleoperated, and how was the learned policy trained and validated?
Evidence to request: Validation process (simulation and real data), update cadence and how a release is rolled back.
- Specifications
Question: Under which conditions was each figure measured, and what is the runtime with your actual payload?
Evidence to request: A spec sheet with test conditions (payload, speed, temperature) and a trial with your configuration.
Visualization methodology
Why we don't use radar charts
The axes of this taxonomy are qualitative and non-linear. A radar chart assigns them arbitrary numbers, and the polygon area — what the eye reads — changes with nothing more than the order of the axes.
Same data, different axis order
The same robot looks “more capable” in one order than in the other.
What we recommend
Parallel coordinates chart
Draws a continuous profile for each model across vertical axes, revealing design and maturity patterns without forcing a single scalar score.
Open the chart →Interactive classification matrix
A two-dimensional interface crossing autonomy level with coordination capability to map system architectures.
Open the matrix →What about the market?
The “tasks × market” automation matrix (use case, robot family, autonomy, sensor fusion and estimated ROI) lives in the ROI calculator.
See the automation matrix →