AI Paper Atlas

Research map / Multi-agent systems and games / Human-AI and goal-conditioned

From Seeing to Experiencing: Scaling Navigation Foundation Models with Reinforcement Learning

Honglin He, Yukai Ma, Brad Squicciarini, Wayne Wu, Bolei Zhou

ICLR 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Vision-language navigation agents

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add RL fine-tuning in simulation after offline pretraining (ablation)S2E-BC · NavBench-GS ObstacleSR: 0.42 -> 0.57Improved
add RL fine-tuning in simulation after offline pretraining (ablation)S2E-BC · NavBench-GS Obstacle + PedestrianSR: 0.40 -> 0.51Improved
scale offline data from 250k to 750k samples (ablation)S2E-BC · success rate (Figure 3a)SR: +2% success rateNo clear effect
use RL instead of SFT for post-training with more training cost (ablation)S2E · URBAN-SIM and NavBench-GSSRImproved
use anchor-guided multimodal GMM instead of single-mode regression (ablation)S2E-BC · NavBench-GS ObstacleSR: 0.33 -> 0.42Improved
replace anchor-guided GMM with discrete anchor classification (ablation)S2E-BC · NavBench-GS ObstacleSR: 0.42 -> 0.37Hurt
replace GMM decoder with diffusion (DiT) decoder (ablation)S2E-BC · NavBench-GS ObstacleSR: 0.42 -> 0.44No clear effect
add offline pretraining before RL (vs PPO from scratch) (ablation)S2E · NavBench-GS ObstacleSR: 0.02 -> 0.57Improved
increase number of anchors from 1 to 64 (ablation)S2E · RECON test setminADE: 0.21 -> 0.09Improved
use 16 anchors instead of 8 (ablation)S2E · RECON test setmAP: 0.62 -> 0.59Hurt
add RL fine-tuning of policy headViNT* · NavBench-GSSR: 0.27 -> 0.39Improved
fine-tune residual attention module instead of action decoder layers with RL (ablation)S2E RL fine-tuning · training stabilityGPU memory: nearly 40GB -> 37GBImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy