Oct 2026· International Conference on Information and Software Technologies· pp. 411-420· 0 citations· 19 references
TL;DR
Long-horizon evaluations across in-distribution and held-out CARLA towns show that the proposed architecture outperforms monolithic and end-to-end MoE baselines, especially under multi-scenario and out-of-distribution conditions.
Abstract
Autonomous driving in urban environments requires policies that handle heterogeneous maneuvers without collapsing into a weak generalist. This paper presents a hierarchical Mixture of Domain Experts framework for deep reinforcement learning in CARLA. We first train Soft Actor-Critic specialists for distinct driving primitives, including intersections, curves, straight roads, and roundabouts. We then freeze these experts and train a gating policy to compose their actions through dense or sparse routing. This separation reduces task interference and enables online skill composition. Long-horizon evaluations across in-distribution and held-out CARLA towns show that the proposed architecture outperforms monolithic and end-to-end MoE baselines, especially under multi-scenario and out-of-distribution conditions.
STRAT, an auxiliary task that trains deep reinforcement learning agents to predict a short textual trace of their own state, solves complex environments where standard RL fails outright, while compacting state representations and preventing rank collapse.
Muhammad Umair Nasir, Alex Vogt, Steven James et al.· 0 citations
MILER, an end-to-end policy framework with zero-shot sim-to-real transfer of both perception and control, is presented and extensively evaluated on a diverse test track comprising numerous challenges.
Thomas Steinecker, Denis Trescher, Alexander Bienemann et al.· 0 citations
This work studies how much closed-loop driving competence a compact multimodal policy can acquire from offline demonstrations in the CARLA simulator, and reports offline metrics and distinguish measured results from qualitative closed-loop observations.
This method formalizes the navigation task as a semiMarkov decision process and constructs a two-layer decision architecture with collaboration between a high-level manager and a low-level worker with collaboration between a high-level manager and a low-level worker.
Qi-Ming Chen· International Conference on...· 0 citations
A navigation policy in which a convolutional network computes a shared spatial field over the map once per timestep and a compact goal-conditioned operator derives agents'movement preferences is proposed.
Jean Seong Bjorn Choe· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.