01
Industrial floors defeat single-sensor autonomy
Long blank aisles, dust, glare, steam and moving equipment make any one perception modality unreliable. The platform needed complementary sensors whose failure conditions do not overlap.
Selected work
Product design case study
Amie
Designing one industrial robot to move material and inspect the plant on the same route.
Amie is a Root Digit Labs programme combining autonomous material movement with continuous condition monitoring. The design treats transport, inspection, safety and fleet learning as one industrial system rather than separate products.
- Programme
- Amie
- Sector
- Industrial robotics and manufacturing
- Engagement
- Product strategy, systems architecture and robotics R&D
- Period
- February 2026–2028 target
The problem
What the programme needed to solve
Factories usually treat material movement and equipment inspection as separate operating systems. One moves loads on demand; the other checks assets periodically, often after early warning signals have already disappeared.
An autonomous mobile robot already travels the aisles, passes the same machines and carries localisation and compute. The product-design problem was to make that movement useful twice—complete the logistics task while collecting condition evidence that maintenance teams can compare over time.
Why it was difficult
Constraints inside the problem
These operating and governance conditions determined what a credible solution had to achieve.
02
Inspection could not compromise the transport task
Monitoring had to happen along the working route without turning every mission into a slow survey. Sensor placement, inference and route planning therefore had to share one operating model.
03
Autonomy required an independent stop path
A sophisticated planner is too large to serve as its own safety case. Motion commands needed a smaller independent monitor and deterministic control path capable of vetoing unsafe behaviour.
04
Plant data had to remain on site
Continuous raw footage can expose processes, people and layouts. The control path and primary inference therefore had to remain at the edge, with fleet learning designed around aggregated updates rather than exported recordings.
The solution
What Root Digit built
Root Digit is designing Amie as an edge-first autonomous mobile platform that performs material handling and condition monitoring on the same route.
The current architecture combines tightly coupled LiDAR-inertial localisation, stereo vision, thermal sensing and structure-borne acoustics with a hardened ROS 2 runtime, a real-time motion layer and an independent safety monitor. A digital twin supports policy training before changes reach physical hardware.
How we did it
Engineering the solution
The workstreams below show how the solution was designed, built and controlled.
01
One operating model for movement and inspection
Route, payload, docking and inspection requirements were modelled together so condition monitoring uses the path the platform already needs to travel instead of creating a second mission system.
- Material-handling missions
- Repeatable inspection routes
- Autonomous docking
- Comparable condition records
02
Multimodal perception and localisation
LiDAR and inertial sensing maintain geometry and motion state, while stereo, thermal and acoustic channels contribute semantic and equipment-condition evidence. The modalities are chosen for different failure modes, not for sensor count.
- LiDAR-inertial odometry
- Stereo semantics
- Radiometric evidence
- Structure-borne acoustics
03
Simulation before hardware release
Navigation, docking and inspection policies are trained across randomised friction, lighting, payload and sensor conditions. Physical release follows validation against the digital twin rather than a single scripted demonstration.
- Domain randomisation
- Payload variation
- Sensor-noise injection
- Scenario-based validation
04
Deterministic motion and independent supervision
A real-time control path is separated from the autonomy stack. An independent runtime monitor checks commanded speed, clearance and deceleration against a defined safety envelope and can veto the planner.
- Real-time control
- Independent motion monitor
- Safety-envelope checks
- Fail-safe stopping
05
Edge-first fleet learning
Primary perception and control remain on the platform. The fleet-learning design shares aggregated model updates so a defect learned at one unit can improve others without exporting raw plant footage.
- On-device inference
- Local control path
- Aggregated learning updates
- No cloud dependency for motion
System design
Architecture and controls
01
Perception path
LiDAR, inertial, stereo, thermal and acoustic evidence is time-aligned into localisation, occupancy and condition models at the edge.
02
Motion path
The autonomy stack proposes behaviour, while the deterministic control layer and independent monitor retain authority to slow or stop the platform.
03
Learning path
Comparable route observations build asset history locally; aggregated updates can improve fleet models without making cloud access part of the control loop.
Delivery sequence
How the programme progressed
Phase 01 · under way
Perception and mapping
Sensor fusion, learned occupancy representation and mapping are being evaluated against recorded industrial environments.
Phase 02 · under way
Autonomy in simulation
Navigation, docking and inspection policies are being trained under domain randomisation and evaluated in the digital twin.
Phase 03 · next
Hardware integration
The deterministic control path, safety monitor and physical sensor payload move onto a reference chassis.
Phase 04 · planned
Supervised pilot
A single unit will run a real route under supervision and be measured against the published design targets.
Outcomes and evidence
What the programme is designed to establish
These are programme design targets and architecture commitments, not measured production or customer-deployment results.
- 24/7
- continuous-operation target
- ± 2 cm
- localisation target
- 4
- fused sensing modalities
- 0
- cloud dependencies in control
Supported by autonomous docking and charge scheduling.
Target accuracy for repeatable routes and asset observations.
LiDAR, stereo vision, thermal sensing and acoustics.
Motion and primary inference are designed to operate on site.
Product design conclusion
Amie turns an unavoidable factory journey into two operating outcomes: material reaches its destination, while the plant builds a comparable condition record under an autonomy architecture designed for industrial safety and data control.
Start a conversation