See the interfaces among data, control, hardware and people.
Robotics technology stack · Layer 06
On-device compute and real-time systems
How deterministic control, local inference, scheduling, thermal design and degraded modes turn compute performance into dependable robot behavior.
Evidence statusEstablished real-time practice; AI workload assurance remains system-specificLast reviewed: 11 August 2026
06
FUURAA thesis
Peak throughput is not real-time performance. A robot needs bounded latency, power and thermal behavior across the complete worst-case workload.
Reading method
Turn a technology label into an engineering chain, testable claims and explicit boundaries.
Translate capability into task, latency, failure and recovery.
Separate standards, independent measurement, research and first-party claims.
State what cannot be inferred from a demo, benchmark or interface.
System breakdown
Four interdependent layers determine whether the technology can enter real work.
Each layer shows its role and the failure signal most worth watching.
Control deadlines
Servo and safety loops require predictable completion before physical dynamics outrun the controller.
Inference pipeline
Preprocessing, memory transfer, model execution and postprocessing all contribute to action latency.
Power and thermal envelope
Sustained frequency, cooling and battery demand determine whether benchmark speed survives a shift.
Degraded operation
Local fallbacks preserve stopping, stabilization and essential sensing when compute or networks fail.
Engineering evaluation
Five checks turn abstract capability into reviewable system evidence.
Record normal performance, failure, recovery and human cost—not only the best-looking result.
- 01
Benchmark tails
Report high-percentile and maximum latency with the full concurrent workload, not only means.
- 02
Run sustained thermal tests
Exercise realistic ambient temperature, enclosure, battery state and workload duration.
- 03
Trace end to end
Measure sensor timestamp to actuator command, including middleware, copies and scheduling.
- 04
Inject resource contention
Stress memory, storage, accelerators and communication while safety-critical loops run.
- 05
Verify offline behavior
Disconnect cloud services and confirm the robot reaches a documented safe, useful or stopped state.
Scope boundaries
State what the evidence supports—and what it does not.
- 01
Hardware acceleration can increase throughput without meeting deterministic deadline requirements.
- 02
A component benchmark excludes middleware, contention, thermal limits and integration overhead unless explicitly included.
- 03
Cloud connectivity should not be assumed for protective stops or stabilization functions.
Sources and evidence status
Keep the source, date, evidence identity and reading boundary visible.
This page prioritises standards bodies, public measurement programmes, official project documentation and original research disclosures.
01
ROS 2 · reviewed 2026-08-11
↗02Real-time programming in ROS 2
Explains deadline, jitter and avoidance of nondeterministic operations in real-time control paths.
ROS · 2021-08-10
↗REP 2008 — ROS 2 Hardware Acceleration Architecture
Proposes vendor-neutral conventions and a benchmark-first method for integrating accelerators.