01 / Problem
Active suspension is an emerging technology developed only for luxury cars — bringing it to public transport requires ground-up innovation
Active suspension is still a frontier technology. What exists today — Bose's research prototypes, ClearMotion's electromagnetic units, Audi's predictive systems — is built exclusively for luxury vehicles at ₹5–15L per system, optimized for high-speed handling on smooth European roads. Nobody has attempted this for public transport. And doing so isn't a matter of making existing systems cheaper — it requires fundamentally new approaches to actuator design, control architecture, and cost engineering.
Active suspension is fundamentally a software problem running on sensors and compute. As AI, edge processors, and MEMS sensors ride exponential cost-performance curves, what is today a ₹10L luxury feature becomes a ₹10k commodity module. The mechanical innovation required is real — but the economics flip because the expensive part (control intelligence) gets cheaper every year. India is the right place to build this: the need is urgent, the scale is massive, and the low-speed operating regime simplifies the engineering.
Indian roads present an extreme input signal: high-amplitude low-frequency events (potholes 50–150mm deep, speed bumps every 200m) superimposed with continuous high-frequency roughness (broken surfaces, gravel, expansion joints). The vibration PSD on a typical Bangalore arterial is 2–10× what existing active systems are designed for.
Low-speed-only operation creates unique engineering advantages:
Relaxed handling constraints. At low speeds, a softer suspension with higher vertical travel provides high comfort with sufficient safety. No compromise for 200 km/h stability. This unlocks actuator and spring designs that luxury cars can't use.
Pre-computable road profiles. Our vehicles repeat routes. Connected fleet vehicles map road surfaces. We can pre-compute suspension commands — where others predict in 10ms real-time, we work with low-cost electronics because we already have the map.
Software-first leverage. Principles from building reliable, performant software systems apply directly to active suspension control. There is significant room to out-innovate incumbents on the software and simulation side.
First-principles 80/20s. The right physics insights and basic modelling may reveal untapped opportunities. We've taken contrarian, simple approaches in software systems by questioning fundamentals — the same method applies to mechanical design.
Target user: a mobile worker commuting 45 minutes on Bangalore roads. Success metric: can you hold a pen steady and write legibly while the vehicle is moving?
02 / Architecture
Three-layer active-passive hybrid
No single actuator spans the full displacement/frequency range. We're designing a layered system — each layer handles a specific regime, the control system coordinates across them.
Layer 1
Large motion
Linear motor — direct electromagnetic actuation for high-amplitude, low-frequency inputs: potholes, speed bumps, road heave. ±50mm stroke, 500N+ force, 0.8ms latency, ~100Hz bandwidth. No gears, no fluid — current in, force out. We're designing this in-house. Interim: ball screw actuator (2.4ms, ~30Hz) while the motor is developed.
1–30 Hz · ±50mm · primary ride control
Layer 2
Fine vibration
VCA (voice coil actuator) swarm — matrix of small actuators beneath the seat, independently controlled. ±5–10mm stroke. Each VCA: 30–50N, collective force scales with count. Exploring moving-magnet configs and multi-stage stacking for stroke extension.
15–100+ Hz · ±5–10mm · localized cancellation
Layer 3
Passive
Magnetic springs (opposing permanent magnets — inverse-cube restoring force) + air springs (tunable pressure) + eddy-current damping (copper near magnets — passive EM dissipation). Zero friction, zero wear, zero power. Works without electronics — physics as fail-safe baseline.
20–200+ Hz · passive · no power · fail-safe
The active actuator is also the damper. No separate shock absorber. EM actuators do skyhook damping — velocity-proportional force referenced to an inertial frame. Energy regenerated as back-EMF. Fail-safe: windings short on power loss → passive braking.
The cost insight: springs and magnets handle static load (free energy storage), opposing magnets provide zero-power negative stiffness — the motor only fights dynamic forces, not gravity. If passive elements are well-tuned, the active actuator does 10% of the work. This is what makes active suspension viable at public transport price points.
03 / Actuator comparison
Why linear motor, and what we use interim
| Mechanism | Latency | Bandwidth | Force | Status |
|---|
| Ball screw | 2.4 ms | ~30 Hz | High | COTS — interim Layer 1 |
| Linear motor | 0.8 ms | ~100 Hz | Med-High | Target — in-house build |
| Hydraulic | 8 ms | ~15 Hz | Very high | Deprioritized |
| VCA (single) | <0.5 ms | 100+ Hz | Low (5–50N) | Layer 2 — in-house |
| VCA (swarm ×20) | <0.5 ms | 100+ Hz | Med (600–1kN) | Layer 2 — experimental |
04 / Open challenges
Unsolved problems — where you come in
Electromagnetic Design
Building a linear motor in-house
Design from first principles: permanent magnet track geometry, coil optimization (fill factor, thermal limits, back-EMF shaping), Halbach arrays for field shaping, air gap tolerance. 2D iteration in FEMM, 3D validation with Elmer FEM (magnetostatics, transient magnetics, coupled thermal). Extract force surface F(position, current) as a surrogate model bridging FEM to real-time control. Target: ±50mm, 100Hz, 500N+.
Simulation Stack
Four-layer simulation pipeline
Layer 1: Electromagnetic FEM — flux density, force curves, saturation, eddy currents. Layer 2: Surrogate extraction — F(x,i) lookup via curve fitting, bridging FEM to real-time. Layer 3: System dynamics in Julia/ModelingToolkit — embed force surface with mechanical dynamics, control laws, symbolic→auto-Jacobian→fast ODE. Layer 4: RL on GPU — NVIDIA Warp for differentiable physics, MuJoCo for rigid body, JAX for end-to-end gradients.
Passive + Active Hybrid
Making the motor do 10% of the work
The cost breakthrough: springs handle static load, opposing magnets (Halbach array) provide zero-power negative stiffness — pre-loading the system so the motor only fights dynamic forces. The spring + magnet negative stiffness system has closed-form dynamics. Simulate and tune this first — the passive layer determines the active motor's force budget and therefore its size, cost, and thermal envelope.
Control Theory
Multi-layer coordinated control
Three actuation layers at different bandwidths and stroke limits need unified real-time coordination. Force allocation across layers, skyhook + feedforward preview + RL-based adaptive tuning. Key advantage: our vehicles repeat routes, connected fleet builds the road map, we pre-compute where others predict in real-time.
RL / Optimization
Non-obvious control policies
Classical control (PID, LQR, H∞) gives a baseline. Nonlinear magnetic springs, stroke saturation, non-stationary inputs may admit better policies. RL in MuJoCo/Warp, sim-to-real transfer. Reward: ISO 2631-weighted RMS acceleration. If the pipeline is end-to-end differentiable (JAX), gradient-based policy search becomes possible over a large non-convex space.
Road Modelling
Indian road surface characterization
No standard dataset captures Indian road severity. Instrument vehicles, drive representative routes, build stochastic road models (PSD-based + transient overlays). Connected fleet vehicles build this dataset over time. Without good road data, we're optimizing for the wrong problem.
Fabrication
Test rig + first physical prototype
Seat mount, road-input simulator (shaker table / eccentric cam), swappable actuator modules, full sensor suite. Commission coil winding, buy NdFeB magnets and linear guide rails, close the loop with a servo drive on a force sensor. Measure F vs. current and compare to FEM. Ground-truth the models with physical data early.
VCA Engineering
Voice coil actuator swarm
Standard VCAs: ±10–20mm stroke. Longer stroke means the coil exits uniform field — force drops nonlinearly. Investigating moving-magnet topology, Halbach arrays, multi-stage stacking. Building VCAs is straightforward — the question is useful stroke range vs. pairing with Layer 1 for large displacements.
05 / Strategy
Four pillars
01
10× simulation and AI-driven design
Simulate before fabricating. Elmer FEM for electromagnetics, Julia/ModelingToolkit for system dynamics, MuJoCo + NVIDIA Warp for RL policy training. Full pipeline: FEM → surrogate model → ODE dynamics → GPU-native RL. All open-source, all GPU-capable. Every physical prototype preceded by hundreds of virtual ones.
02
Creative hardware prototyping
Desktop test rigs before vehicle installs. Commission coil winding, close the loop with a servo drive and force sensor early. Instrumented testing with data feedback into simulation. Build small, measure everything, iterate in days not months.
03
Layer existing tech, invent what's missing
Linear motors, VCAs, magnetic springs, eddy-current dampers, air springs — all exist individually. Innovation: the layered architecture where passive elements carry 90% of the load. Where COTS fails, build in-house. Modular: features toggle for cost — luxury seat to public transit.
04
Convert hardware → software problems
Active suspension = control on sensors + compute. Moore's law, better MEMS, RL advances — all compound over time. Connected fleet builds road maps that enable pre-computation with low-cost electronics. Push mechanical complexity into software where exponential trends work in our favor.
06 / Who we need
Versatile people who get things done across domains
This is early-stage invention. There is no playbook. You'll derive transfer functions in the morning, visit a coil winding shop in the afternoon, and discuss magnet geometry with a professor in the evening. We need people who move fluidly between theory and practice — who are as comfortable with a lathe as with a Lagrangian, and who will travel to meet professors, factory workers, and component suppliers to make things happen.
We don't care about titles or credentials. We care about whether you can make something work that didn't exist before — and whether the reason you want to do it is because 180 million people ride broken roads every day and you think that's a problem worth solving with your hands and your mind.
Hands-on work
- Machining — lathe, mill, drill press
- Welding — TIG, MIG, stick (steel, aluminum)
- Motor winding, coil fabrication, magnet assembly
- Soldering, harness building, PCB rework
- Sensor integration — accelerometers, IMUs, encoders
- Test rig assembly and instrumentation
- Visiting shops, sourcing components, getting quotes
Theoretical work
- Dynamics — Lagrangian mechanics, multi-body systems
- Control — PID, LQR, H∞, skyhook, adaptive
- Electromagnetics — Lorentz force, magnetic circuits, FEA
- RL / optimization — policy gradient, sim-to-real
- Signal processing — PSD, transfer functions, filtering
- Simulation — Julia, MuJoCo, Python scientific stack
- First-principles physics modelling and intuition
"Use inventive technology and AI to do larger good for the public — build transportation so comfortable and affordable that a commuter can work undisturbed on the worst road surface in the world."
The purpose