Intelligence, made to fit.
LINKU integrates AI and control into compact machines — designed around the task, able to run on-site, and tunable as conditions evolve.
Designed for the task
We design around a defined task. From sensing to model, each layer can be tuned to the setting, its constraints, and changes over time.
Can run on-site
Computation and decisions can happen on the device, avoiding a required cloud round trip. Depending on deployment, the system can continue offline and keep data local — useful where latency, connectivity, or data handling matters.
Value by design
Hardware specifications define the boundary, not the outcome. Co-design coordinates models, control, sensing, and hardware to approach the system's practical limits. We focus on useful task performance within those constraints.
Designed to adapt
Where a deployment supports data collection and updates, field data can inform later tuning, model updates, or control adjustments. Improvement is an engineering process, not an automatic promise.
Same simulated hardware, two control strategies.
Both sides simulate the same plant, target, disturbance, and noise. The left uses a competent fixed-gain PID; the right uses model-based control built from public best practices on the same simulated hardware. This illustrates one control-layer contribution to software–hardware co-design, not physical hardware validation. Faster settling can support higher cycle rates; lower error can reduce the margin required from mechanics and sensing.
Simulated illustration, not measured data · Both sides share one target, one disturbance, one noise sequence · Honors prefers-reduced-motion; pauses off-screen
Move the cursor to move the target · Click to hit both sides with the same disturbance · The thin line below traces error over time — flatter is better
Tap the rings to set the target · Tap Disturb to hit both sides at once · The thin line below traces error over time — flatter is better
For engineers: model & parameters
Sensing (identical) encoder quantization 1.5° (≈240 CPR), Gaussian noise σ = 0.5°; both sides consume the same noise sequence (common random numbers)
Common practice (a competent baseline) fixed-gain PID: τ = 20·e + ∫12·e − 5.8·ω̂f (ζ ≈ 0.65); velocity by 6-tick encoder differencing + first-order low-pass (dirty derivative, ω ≈ 20 rad/s); anti-windup by conditional integration (paused while the output saturates); raw step commands — no trajectory, no model feedforward, no observer, no latency compensation
Model-based control (public best practice, deployed correctly) steady-state Kalman (α–β) estimate + one-tick forward prediction; near-time-optimal velocity governor (PTOS: v ≤ √(2a·e) far out, linear terminal zone near; a ≤ 0.8·τmax/J keeps feedback headroom); model feedforward τff = J·a_ref + b·ω_ref; momentum disturbance observer (ω ≈ 12 rad/s) estimates and cancels external torque; feedback τfb = 40·(θref−θ̂) + 11·(ωref−ω̂), ζ ≈ 0.87
Disturbance click = equal impulse torque to both sides; ambient torque noise is continuous and shared
Statistics settling = first entry into the ±2° band held for 0.3 s (same definition both sides; the full error history is the thin trace below)
Scope everything on the right is public, textbook-grade technique; this simulation isolates one control-layer contribution to co-design and is not measured hardware evidence
Get in touch.
Have a task you'd like to make smarter, or a collaboration in mind? We'd love to hear from you.