Bare coding agent hits 78% on R2R-CE navigation with zero training, no mapping or memory

jiqizhixin · x · 2026-09-06

University of Adelaide's MIP (Minimal-Interface zero-shot agents) drops a general-purpose coding agent straight into a physical environment—no mapping module, no memory, no waypoint predictor, no domain-specific code or navigation training. The agent reads raw observations and outputs raw actions via a minimal interface, relying purely on general reasoning.

Result: 78% success rate on the classic vision-and-language navigation benchmark R2R-CE, matching or beating recently released industrial-scale navigation models trained on massive specialized data (humans: 94%). The finding raises the question of whether industrial navigation models have been solving the wrong problem.

Original post →

More from Embodied

Embodied channel →