Inspired by LLMs he announces a post-training era for AS agents: collect and generate data (with 3DGS + behavior world models for trajectories or soon video world models) and train with RL #agents4ad #cvpr2026
.@francislee2020 is taking the stage at #agents4ad Protip: make sure to set a google scholar alert for his papers #cvpr2026
Key result: post-training gains from 50k scenes would need 14x more data for behavior cloning pretraining #agents4ad #cvpr2026
Another key brick is the behavior world models thar generates agent trajectories and reactions and communicated with the 3DGS simulator for RL training: beautiful execution! #agents4ad #cvpr2026
Inspired by LLMs he announces a post-training era for AS agents: collect and generate data (with 3DGS + behavior world models for trajectories or soon video world models) and train with RL #agents4ad #cvpr2026
.@francislee2020 is taking the stage at #agents4ad Protip: make sure to set a google scholar alert for his papers #cvpr2026
Users praise the behavior world models for autonomous driving agents at CVPR for their beautiful execution in generating trajectories and reactions integrated with 3DGS simulators for RL training.
Key result: post-training gains from 50k scenes would need 14x more data for behavior cloning pretraining #agents4ad #cvpr2026
Another key brick is the behavior world models thar generates agent trajectories and reactions and communicated with the 3DGS simulator for RL training: beautiful execution! #agents4ad #cvpr2026
One of the key ingredients in their data engine, MTGS, still did not get accepted anywhere so far in spite of its great utility: MTGS: Multi-Traversal Gaussian Splatting https://arxiv.org/abs/2503.12552 #afents4ad #cvpr2026