Perception Engineer

Advanced Terrain-Aware Perception

• Implement material-specific semantic layers (soil, sand, gravel, bed-rock, tall grass,

shallow water, snow, ice) that update every frame and age-out when confidence degrades.

• Derive surface normal, roughness, cohesion and expected sink-age from texture

statistics + monocular depth cues.

• Compute instantaneous grade (), side-slope, slip angle and flag combinations that

exceed rollover or traction thresholds.

• Detects ruts, berms, holes and negative obstacles by fusing long-baseline disparity

with shading gradients and shadow parallax.

• Class-specific hazard masks (e.g. sage-brush vs. tumbleweed vs. basalt outcrop) that

feed directly into the traversability cost map.

WheelTerrain Interaction Estimation

• Fuse visual shear-flow with wheel tachometry to produce online slip-ratio and effective

(friction coefficient) per wheel.

• Predict sink-in depth (sand / snow) from colour histogram shifts and wheel torque

spikes.

• Raise soft faults when wheelground contact probability less than0.3 (airborne cresting, pothole

drop-in, obstacle hang-up).

Real-Time Route Feasibility & Micro-Planner Hooks

• Transform pixel-level terrain classes into steer-able corridors and hazard polygons with

less than 25 ms latency.

• Prioritise dynamic re-weighting of cost metrics (e.g. favour firmer gravel over muddy

shortcut when rainfall detected).

Extreme Environment Hardening

• Design visibility-fallback perception:

Short-range active illuminator gating for dust / snow glare.

Polarised camera processing for water-surface discrimination.

Thermal fusion when visible light SNR less than 6 dB.

• Implement turbulence-robust line detection to keep horizon and vanishing-point

estimates stable during heat shimmer.

Self-Consistency Audits

• Every 100 ms run terrain confusion matrixdrop or re-sample tiles whose posterior

variance greater than 2_thresh.

• Maintain loop-closure integrity score for vision odometry; trigger re-initialisation if drift

greater than 2 % of path length within 30 s.

Additional Required Expertise

Area Advanced Expectation

Photometric Stereo Reconstruct micro-surface undulations for traction estimation

without LIDAR.

Rolling/Center-of-Mas

s Stability

Predict rollover using tilt-vector + speed + suspension travel in

real time.

Dust & Snow

Morphology

Implement particle-size adaptive temporal filtering to separate

airborne particulates from solid obstacles.

Realtime SIMD / DSP Hand-vectorise convolutional kernels for ARM NEON or x86

AVX-512 where no GPU is present.

Failure Injection Design unit tests that inject random camera blackout, saturated

highlights, frame drops, inertial spikesand verify graceful

degradation.

Bonus Capabilities

• Built perception for Dakar-class rally bots, planetary rovers or Online Desert

Challenge vehicles.

• Published work on vision-based slip prediction or terrain type transfer learning

in CVPR, ICRA or RA-LAT.

• Experience creating synthetic terrain datasets with physically-based rendering for

rare surfaces (lava crust, talc powder dunes).

• Knowledge of acoustic ground contact sensing and correlating it with visual

classification for redundancy.