- The central framing emphasizes a single pretrained VLM decoder and minimal architectural deviation. However, the method adds structured-query self-attention, separate task-specific FFNs, task heads, 3D positional embeddings, and a multi-stage adaptation protocol. These are reasonable design choices, but they are substantial task-specific components. Thus, the practical contribution is better characterized as a shared-attention multitask decoder with task-specific transformations, rather than a largely unmodified unified VLM decoder.
The paper should explicitly quantify the additional parameters, FLOPs, and training cost introduced by query self-attention and the per-task FFNs. It should also clarify exactly which modules are shared, copied, frozen, LoRA-adapted, or independently learned at each stage.
- The diagnostic experiment in Table 1 is suggestive, but not sufficient to support the stronger claim that pretrained VLM causal attention is the key transferable ingredient enabling OneDrive. The study compares initialization choices in a parallel-decoder setting, but the paper would benefit from controlled ablations in the actual OneDrive architecture, such as:
random initialization of causal attention versus pretrained causal attention;
frozen pretrained attention versus finetuned attention;
replacing causal attention with standard task-specific cross-attention;
retaining pretrained FFNs versus task-specific FFNs under matched optimization/training budgets;
removing perception-query self-attention;
using separate decoders with comparable parameter count and supervision.
Without these controls, it is difficult to isolate whether gains arise primarily from pretrained causal attention, the multi-stage curriculum, task-specific FFNs, query ordering, added detection supervision, or their combination.
- The nuScenes comparison includes text-based and action-based methods with potentially different inputs, pretraining sources, language annotations, ego-state usage, model sizes, and evaluation conventions. Although the paper distinguishes the two categories, the overall “state-of-the-art” conclusion may not be directly comparable across all rows.
Likewise, NAVSIM comparisons may involve substantially different visual backbones, training data, auxiliary supervision, and sensor modalities. The paper states that OneDrive uses no additional autonomous-driving data for pretraining, but it would be helpful to precisely specify:
all data used for each training stage;
whether OmniDrive QA annotations or derived language annotations are used;
whether the InternVL3 pretraining data overlap with relevant driving/image domains;
model parameter counts for all OneDrive variants;
input temporal context, image resolution, camera configuration, and ego-state inputs;
the exact NAVSIM training/evaluation protocol and whether all compared methods use identical data splits.
Further, all main results should include variation across multiple random seeds or confidence intervals. A difference of 0.05 PDMS or small collision-rate changes can be difficult to interpret without variance estimates.
- The evidence for preserved language capability is limited
A main claim is that OneDrive preserves multimodal/text generation capability while jointly learning structured driving tasks. However, the language evaluation mainly compares text-conditioned trajectory prediction L2 against OmniDrive. This does not directly evaluate textual generation quality, factual grounding, or whether the language output remains causally useful after joint finetuning.
I encourage the authors to add:
standard driving VQA/text-generation metrics, where applicable;
qualitative examples with generated scene descriptions, reasoning, and planned trajectories;
human or automatic grounding checks showing whether generated text correctly reflects detected objects, lanes, and ego actions;
interventions that alter language instructions while holding images fixed, and measure the resulting behavioral changes;
tests for whether text supervision improves behavior under language-dependent or ambiguous scenarios, not only aggregate planning metrics.
This is particularly important because Table 5 shows only modest aggregate gains from text supervision.
- The 40% latency reduction is compelling for NAVSIM relative to the reported ReCogDrive runtime. However, the nuScenes latency table explicitly notes that attention kernels are not matched: OneDrive uses FlashAttention while several comparison values do not. Therefore, the nuScenes efficiency comparison does not isolate architecture-level efficiency.
Moreover, the complete nuScenes configuration is reported at 513 ms/frame, which raises questions about real-time deployability for a surround-view planning system. The paper should report:
matched-hardware and matched-kernel latency;
throughput and memory usage;
latency breakdown for ViT encoding, decoder execution, perception heads, and planning;
latency as a function of number of cameras, number of image tokens, decoder depth, and output queries;
performance–latency trade-offs for the shallow/truncated mode relative to the full model.
- The evaluation is restricted to nuScenes and NAVSIM, which the paper itself acknowledges. For a VLA-based framework claiming unified multimodal reasoning and improved planning, stronger tests of generalization would substantially strengthen the paper:
long-tail and safety-critical subsets;
adverse weather, nighttime, occlusions, and rare interactions;
cross-city or cross-domain transfer;
command ambiguity and counterfactual language instructions;
out-of-distribution camera conditions;
closed-loop scenario-category breakdowns rather than only aggregate PDMS.
The limitations section notes ambiguous/mismatched commands, but this should be elevated into a systematic evaluation rather than left primarily as future work.