Reproduce “GSDrive: Reinforcing Driving Policies by Multi-mode Future Trajectory Probing with 3D Gaussian Splatting Environment” on OpenResearch

About this paper

Ziang Guo, Chen Min, Xuefeng Zhang, Yixiao Zhou, Shuo Wang, Sifa Zheng, Dzmitry Tsetserukou, Zufeng Zhang

Researchers developed GSDrive, a framework for end-to-end driving policies that leverages a differentiable 3D Gaussian Splatting environment to generate dense, future-aware reward signals. This approach resulted in superior driving performance, achieving the lowest collision rate of 0.11 and the highest episode reward of 52.97 in closed-loop evaluations.

Key insights

• Dense, future-aware reward shaping, derived from probing multi-mode trajectories in a 3DGS environment, effectively addresses the sparse reward problem in RL, leading to more stable and higher-performing policies.
• The cyclic hybrid IL-RL training loop successfully combines the strengths of both paradigms, using IL to provide structured priors and RL to enable continuous policy improvement and generalization beyond expert demonstrations.
• Dynamically weighting trajectory-constructed action logits with residual action logits allows for flexible policy control, balancing the influence of learned priors with adaptability to novel situations.

Summary by alphaXiv.