- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The contribution of this work can be divided into three parts:
- analysis of the mode diversity under different imitation training recipes: number of epochs, number of scene tokens, output representation (velocity vs. trajectory) and denoising path (linear vs. GVP).
- A shared noise vocabulary selected offline from Gaussian noise to maximize decoded-trajectory coverage and quality across training scenes.
- Cross-scene RL: suppress mode collapse in GRPO by grouping the scenes instead of trajectories. Results are validated on three benchmarks demonstrating advanced performance. However, driving performance relies on not only trajectory generation, but also heavily relies on selection, which is underdiscussed in this work.
Critical strengths
The proposed cross-scene RL is novel, and the noise vocabulary is a simple, model-agnostic alternative to handcrafted trajectory anchors. The method shows strong performance on three benchmarks.
Critical weaknesses
Regarding the pretraining recipe, the work shows some empirical results but does not analyze the observations or provide insights into it. Specifically:
- Is diversity equivalent to quality? The focus of this work is preserving the diversity during pre- and posttraining, while the diversity is not properly measured. The work treats best-of-K performance (line213) as the diversity indicator, but lacks justification.
- v-pred and x-pred share the same optimal solution and can be converted into each other via Eq. 8. To my understanding, v-pred implicitly upweights low-noise flow timesteps relative to x-pred, by a factor that diverges as from t to 1 under GVP. A deeper analysis of why this leads to higher diversity is needed.
On the cross-scene RL side, the idea is novel but the effect seems to be minor on NAVSIM and HUGSIM.
- Details of the trajectory selection is missing. How do you train the critic ensemble? How do you weight the predicted subscores to select trajectory? Simply switching the backbone shows more significant improvements (more than 1 PDMS, Tab 6) than the proposed RL method (0.4 PDMS, Tab1).
- The performance gains seem to be mainly from the better backbone. The SotA method DrivoR uses DINOv2-S as the backbone with a single critic , achieving 93.1 PDMS, while this work achieves 92.55 PDMS with the same backbone with a single critic (Tab. 6), behind DrivoR.
Questions for authors
Please provide details about the critic training and inference
- how is it trained, what is the supervision targets, is it jointly trained with the trajectory generator, do the critic and generator share the backbone...
- what's the difference of each critic in the ensemble? training seed, or something else?
- how are the predicted scores used? How is the trajectory selected?
Can you report results where both the generator and the critic use a backbone comparable to DrivoR (DINOv2-S), so that gains from the proposed method can be separated from gains due to a stronger backbone?
Can you provide the diversity measurement and its correlation to the best-of-K performance?