All three preprints address the same underlying problem: autonomous driving policies trained in simulation or on logged data often fail when faced with the open world. MILER argues that reinforcement learning has been held back by the sim-to-real gap, especially in unstructured environments, and proposes a semantic mid-level representation to make learned policies more transferable. REARL, by contrast, targets the simulator itself, using real traffic data and large language models to generate closed-loop scenarios that better reflect real-world dynamics. OPTED takes a post-training route, noting that open-loop pre-training alone yields diminishing returns and that end-to-end policies need on-policy fine-tuning.
The papers differ in where they intervene. MILER changes what the policy perceives; REARL changes what the simulator generates; OPTED changes how the policy is trained after pre-training. They also differ in emphasis: MILER targets unstructured driving, REARL targets traffic complexity, and OPTED targets the open-loop-to-closed-loop mismatch. None of the three claims to have solved the problem outright, and the abstracts do not report direct comparisons against one another.
Read together, they suggest a field converging on the same diagnosis—simulation and pre-training are not enough—while still exploring distinct remedies. That divergence is healthy, but it also means the practical value of each approach will depend on head-to-head evaluation in real-world or high-fidelity closed-loop settings, which none of the abstracts provide.