Micro Quest Book - Fall 2026 (F26)

The Great Objective: Competition Ready

This term, Micro aims to improve competition of our MVP stack, optimize existing algorithms and start work on micro-BEV.

[Brief overview of what will be accomplished this term, organized into categories.]

Hardware
- Integrate the Orin Nx
- Car Maintenance
Software
- MPC Controller
- GPU-Accelerated Particle Filter
- Object Detection pipeline
- semantic segmentation pipeline

Score	Criteria
7/7	Mounted to the car with all packages installed, sensors hooked up and software stack running
4/7	all packages and stack installed and setup
0/7	No progress made / was unable to buy NX

Minimum Requirements:

Score	Criteria
3/3	New platform plate mounted to the car, replacing the broken one
2/3	New platform plate printed
0/3	No progress made

Minimum Requirements: New platform plate printed for score of 2/3

Score	Criteria
10/10	Deployed on car, real-time validated, tuned race-ready
8/10	Runs on the car in closed loop, tracks reasonably, but not yet race-tuned / edge cases (solver failure, infeasibility) not fully handled
6/10	Closed-loop in sim with real state estimation + reference generator integrated; warm-starting and fallback logic in place
5/10	Closed-loop in sim, tracks reference, respects constraints, stable
3/10	Full problem formulated in the library API, solves repeatedly, but loop/reference not yet wired
2/10	Correct single-step solution from a valid formulation
0/10	No functional MPC

Minimum Requirements: Full problem formulated in the library API, solves repeatedly, but loop/reference not yet wired

Score	Criteria
10/10	CUDA PF deployed on Jetson, hits ≥50 Hz at target particle count, Jetson-optimized (unified/pinned memory, fused kernels, FP16 where valid, power/clock profile tuned), localization accuracy at parity with CPU reference, running live in the loop
8/10	CUDA PF runs on Jetson in the localization loop at real-time rates but below the 50 Hz target, or hits 50 Hz on desktop GPU but not yet Jetson-optimized; accuracy verified against CPU
6/10	Full CUDA acceleration (sample, weight, resample on GPU) working with a measured, meaningful speedup over the Python GPU baseline; accuracy matches CPU version; not yet ported/tuned for Jetson
5/10	Partial GPU acceleration — hottest stages (e.g. weighting) moved to CUDA, rest on CPU; correctness held, speedup demonstrated
4/10	CPU PF profiled and benchmarked; bottlenecks identified and GPU port planned; CUDA scaffolding/build set up
3/10	CPU PF complete and running in the localization loop at parity with reference accuracy, serving as the correctness baseline
2/10	CPU PF implemented (predict, update, resample) and tracking correctly on recorded/offline data
0/10	No functional particle filter

Minimum Requirements: CPU PF complete and running in the localization loop at parity with reference accuracy, serving as the correctness baseline

Score	Criteria
10/10	Deployed on car, real-time on live feed, latency-optimized (TensorRT/quantization), detections tracked with stable IDs, robust across lighting/weather
8/10	Real-time on live camera feed, all target classes, integrated into stack with defined output contract (bbox + class + confidence + timestamp), but not latency-optimized and no cross-frame tracking
6/10	Runs online on live feed at usable rate, all target classes, output published to the stack but contract/format not finalized
5/10	Runs offline on recorded data, detects all target classes reliably, output visualized
3/10	Runs offline, detects a subset of classes, visualized but accuracy inconsistent
2/10	Model loads and produces detections on a single frame/image
0/10	No functional detection pipeline

Minimum Requirements: Model loads and produces detections on a single frame/image

Score	Criteria
10/10	Deployed on car, real-time free-space segmentation, projection validated against ground truth, temporally consistent masks, graceful under calibration drift
8/10	Real-time free-space segmentation with mask projected into ground/vehicle frame producing a usable free-space region, integrated into stack, but not temporally smoothed or ground-truth-validated
6/10	Real-time segmentation on live feed, projection implemented (IPM/calibration-based) but not yet validated for metric accuracy
5/10	Free-space masks produced online at usable rate, projection prototyped offline
3/10	Segmentation model produces free-space masks offline on recorded frames, no projection
2/10	Model loads and produces a segmentation mask on a single frame
0/10	No functional segmentation pipeline

Minimum Requirements: Model loads and produces a segmentation mask on a single frame

Quest Name	Description	Score
Nvidia Jetson Nx Setup	Mount the Jetson Nx to the car with all packages installed, sensors hooked up and the software stack running
Car Maintenance	Print a new platform plate and mount it to the car, replacing the broken one

Quest Name	Description	Score
MPC Trajectory-Tracking Controller	Build an MPC controller that tracks the reference trajectory in closed loop, deployed on the car and tuned race-ready
GPU-Accelerated Particle Filter	Port the particle filter to CUDA and optimize it for the Jetson, hitting ≥50 Hz with accuracy at parity with the CPU reference
Object Detection Pipeline	Build a real-time, latency-optimized object detection pipeline on the live camera feed with tracked detections integrated into the stack
Semantic Segmentation Pipeline (Free Space) + Camera Projection	Build a real-time free-space segmentation pipeline with the mask projected into the vehicle frame and validated against ground truth