IVALab
Undergraduate Research Assistant
Prof. Patricio Vela
Jun 2026 – present
-
Planner benchmark harness
A seeded evaluation harness that benchmarks 8 autonomous-driving planners on crash, merge, and runtime metrics. Batch runs fan out across 8 workers with ProcessPoolExecutor.
-
Real-time gap planner + CBF-QP safety filter
Ported an offline MATLAB gap-based planner to a closed-loop Python controller for HighwayEnv, and added a CBF-QP safety filter that keeps lane-change trajectories physically feasible so merges stay continuous. It runs a 12 s space-time pipeline in under 4 ms per decision.
-
Acceleration-aware behavior prediction
Prototyped acceleration-aware behavior prediction for surrounding vehicles, measured across 500 paired episodes (p < 0.0001).
-
One scenario format for two languages
A shared JSON scenario format lets the MATLAB and Python planners run identical test cases. Running them side by side exposed a bug in the Python port that dropped valid merge options below the theoretical n+1 bound, which I fixed.
scenario.json (shape, illustrative) { "seed": …, "lanes": …, "ego": { "x": …, "v": … }, "vehicles": [ { "lane": …, "x": …, "v": …, "a": … } ] } MATLAB --+ +--> same cases --> diff --> n+1 bound bug Python --+



