IntentionNav - Affordance Intents: leaderboard

IntentionNav (2026) tests intent-driven object navigation: an embodied agent receives an indirect human intent instead of a target name and must infer the object, find it in an Isaac Sim scene and stop near it, with the evaluated VLM inside a fixed reference agent. This board is the success rate in percent on intents where only what the target does or looks like implies the target; higher is better.

Metric: Success rate (%) within 2.0 m on the affordance intents (only what the target does or looks like is given), over IntentionNav's 2,000 episodes (500 implicit-intent instructions in four styles, 176 Isaac Sim scenes, 64 target categories, target name withheld), the VLM driving the benchmark's fixed reference active-navigation agent (RGB-D, open-vocabulary detection, map memory, 30-step budget); higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.6 Plus20.8#238
2Gemini 3.1 Flash Lite18.2#371
3GPT-5.416.5#135

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=intentionnav-affordance-intents · How It Works · Data refreshed daily, snapshot 2026-11-21.