New spatial planning benchmark: GPT-6 and Claude Opus 5.5 fail parking route tasks

songhan_mit · x · 2026-11-22

Researcher Jinglun Zhao built a spatial planning evaluation for LLM agents using parking as a testbed, covering 6 layouts (offset-entry, dead-end, T/L-shaped aisles, angled bays, curb extensions) scored on distance, gear changes, and clearance. Key findings: manipulation tasks are nearing saturation, but spatial planning still separates models sharply — GPT-6-Astra failed to park on the first two layouts every time, Claude Opus 5.5 never found a valid route on either, and among valid routes quality varied wildly (scores 78.8 vs 11.0, worst clearance 5.90 cm). Routes were validated with full-body collision checks from recorded trajectories.

Original post →

More from Models

Models channel →