acceptodds
Withdrawn from ICLR 2027

AffordDrive3D: Affordance-Aware World-Action Modeling with Spatial Understanding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World-action models have recently improved autonomous driving by jointly learning future scene prediction and trajectory generation. Most existing approaches model the future primarily through RGB appearance, and recent works have begun to incorporate geometric prediction to improve spatial understanding. However, dense geometry describes the spatial layout of the entire scene without indicating which parts are most relevant to the ego vehicle’s action. For driving, the model must also identify and anticipate where it can safely move and which regions may pose collision risks. When action-relevant regions are jointly learned with future geometry, the policy can capture both what matters for driving and the corresponding geometric information. We therefore propose AffordDrive3D, an affordance- and geometry-aware world-action model that jointly learns future action-relevant regions and spatial structure. In order to capture the scene semantics and driving context needed for driving affordance prediction, we build AffordDrive3D on a VLM backbone to forecast drivable areas and collision-critical regions that directly affect ego motion, while predicting future geometry from RGB world-model latents. Experiments on NAVSIM show that AffordDrive3D achieves strong performance and demonstrate the effectiveness of jointly modeling future affordance and geometry.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Vision-Language-Action Models, World-Action Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AarROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AffordDrive3D, an affordance- and geometry-aware World-Action Model for end-to-end autonomous driving. Built upon a shared vision-language model (VLM) backbone, the proposed framework jointly models future RGB frames, depth, drivable regions, collision-critical regions, and ego-vehicle trajectories. Specifically, the authors leverage predicted RGB latents to facilitate future depth prediction and introduce Temporal Affordance Forecasting to enable the model to learn action-relevant spatial information in both current and future scenes. The paper reports 91.3 PDMS on NAVSIM v1 and 89.9 EPDMS on NAVSIM v2, supported by ablation studies and qualitative visualizations.

Critical strengths

1.Well-motivated research direction. Rather than predicting only future RGB observations or geometric structures, the paper additionally models spatial regions relevant to ego-vehicle planning. This makes the supervision objectives of the world model more closely aligned with downstream driving decisions.

2.Task-relevant affordance label design. The construction of collision-critical regions through collision analysis of candidate trajectories, rather than directly relying on conventional object-category labels, is a meaningful design choice that enhances the task relevance of the supervision signals (Appendix A.1).

3.Relatively comprehensive experimental evaluation. Table 3 demonstrates the positive contributions of different auxiliary branches to planning performance, while Tables 4–5 further investigate temporal prediction capabilities and VLM representations, providing useful experimental observations.

Critical weaknesses

1.Lack of a direct ablation between current-frame and temporal affordance supervision. Table 3 demonstrates that adding affordance supervision is beneficial, but it does not establish whether the improvement comes from spatial perception of the current frame or from forecasting future affordances. Consequently, the necessity of the proposed Temporal Affordance Forecasting mechanism has not been sufficiently validated.

2.The paper does not compare independent depth prediction against RGB-conditioned depth prediction, nor does it report quantitative accuracy for future geometric predictions. This makes it difficult to assess the contribution of RGB conditioning to geometry prediction and downstream planning performance.

3.The full model in Table 3 achieves 89.6 PDMS, whereas the final reported result reaches 91.3 PDMS. However, component-level ablations under identical RL training conditions are missing, as are experiments examining combinations such as Affordance + Geometry. This makes it difficult to determine the independent contributions and complementarity of individual components.

4.The SOTA claim and baseline comparisons require clarification. Table 1 does not include the published DriveSuprim method, which reports 93.5 PDMS on NAVSIM v1. Given that DriveSuprim reports 93.5 PDMS on NAVSIM v1, the paper should clarify whether the evaluation settings are directly comparable and specify the scope of its SOTA claim.

Minor Weaknesses

1.The experiments are restricted to NAVSIM. Although results are reported on both v1 and v2, these are not independent datasets. The generalization capability of the proposed method under different data distributions and more challenging driving environments remains unclear.

2.Lack of performance stability analysis. Some improvements over competing methods are smaller than 1 PDMS point, yet the paper does not report variations across random seeds or provide statistical analyses.

3.Although the auxiliary branches are not used during inference, the final system still employs a relatively large VLM and Action DiT. The paper does not report inference latency, memory requirements, or computational cost comparisons with other approaches.

4.Limited analysis of long-horizon affordance prediction. Table 4 shows that collision-critical AP decreases to 0.732 at the 2–4 s prediction horizon. However, the paper does not further investigate the associated failure cases or the relationship between prediction errors and planning performance.

Questions for authors

1.Table 3 shows that affordance supervision improves planning performance, but does not separate current-frame supervision from future prediction. How does the model perform when only current-frame affordance supervision is used?

2.In Appendix A.1, the collision-critical object set is determined at the initial timestep and shared across future frames. How does this account for objects that become collision-critical at later timesteps?

3.Is conditioning depth prediction on RGB latents necessary? A comparison with independent depth prediction would help clarify its contribution to planning performance.

4.The full model in Table 3 achieves 89.6 PDMS, while the final result reaches 91.3 PDMS after RL training. Can the authors provide ablations under the same RL setting to clarify the contributions of the individual components?

5.Table 1 does not include DriveSuprim, which reports 93.5 PDMS on NAVSIM v1. Is there a difference in evaluation settings that prevents a direct comparison? The scope of the SOTA claim should be clarified.

2 comments

We thank the reviewer for the comments.

W1&Q1. Current-frame vs. temporal affordance supervision

We ablate the temporal component under the affordance–action setting (SFT):

AffordanceNCDACTTCEPPDMS
None (action only)97.996.494.382.688.1
Current only98.196.994.383.788.8
Current + future98.397.195.183.589.2

Current-frame supervision improves spatial perception (+0.7 PDMS), while future affordance forecasting brings further gains, most notably in TTC (94.3→95.1), which is unchanged with current-frame supervision alone. This indicates that forecasting future drivable and collision-critical regions helps the model anticipate upcoming conflicts, validating the necessity of Temporal Affordance Forecasting.

W2&Q3. RGB-conditioned vs. independent depth prediction

(a) Effect of conditioning on predicted RGB. Under the geometry–action setting (SFT), we compare conditioning depth on the predicted future RGB latents against copying the current-frame RGB latent to all future latent groups. In both variants, no RGB loss gradient reaches the VLM, so the only difference is the conditioning signal (depth metrics calculated on the front view against DA3 pseudo-GT):

RGB latentTTCPDMSAbsRel +2 sδ1.25 +2 sAbsRel +4 sδ1.25 +4 s
Current94.688.80.2030.7670.2480.714
Predicted (ours)95.289.10.1950.7880.2400.744

Conditioning on predicted future RGB improves both future depth accuracy and planning, especially TTC.

(b) Future depth quality. Our full model substantially outperforms a copy-last baseline that reuses the current-frame DA3 depth for future frames:

MethodAbsRel +2 sδ1.25 +2 sAbsRel +4 sδ1.25 +4 s
Copy-last (DA3)0.3290.7060.4230.631
Ours (full model)0.1910.7900.2350.744

W3&Q4. Component ablations under the RL setting

We apply the same RL training to all settings in Table 3:

SettingSFT PDMSRL PDMS
Action only88.190.2
+ Aff.89.291.0
+ RGB88.990.7
+ Geo.89.190.8
Full model89.691.3

The gains from each component persist after RL. Since the VLM is frozen during RL and only the action DiT is updated, these gains come from the shared representations learned with auxiliary supervision during SFT, which continue to benefit planning after RL. The full model outperforms every single-component variant, indicating that the components are complementary.

W4&Q5. Comparison with DriveSuprim and scope of SOTA claim

DriveSuprim is included in our NAVSIM v2 comparison (Table 2), which uses the extended EPDMS metric with additional rule-compliance and comfort criteria and thus provides a more comprehensive evaluation. Under this setting, AffordDrive3D outperforms DriveSuprim, with higher lane keeping (97.4 vs. 97.0), consistent with our drivable-area affordance that guides the model to drive within lanes in the legal travel direction.

Q2. Collision-critical objects emerging at later timesteps

Objects that become critical at later timesteps are already captured. We roll out each candidate trajectory over the full 4 s horizon, with other agents replaying their logged future motion, and mark every object involved in an at-fault collision at any point during the rollout (e.g., a vehicle cutting in at 3 s). The only objects excluded are those not present in the initial frame: since they cannot be inferred from the input observation, labeling them would introduce supervision the model cannot learn from and act as label noise. This also keeps the labels aligned with the planning decision at t0, from which the trajectory is generated.

Minor W1. Generalization beyond NAVSIM

We additionally evaluate AffordDrive3D in closed loop on NavSafe-∞ [1], a photorealistic benchmark of 280 event-centered scenarios spanning 28 safety-critical event types, built on 3DGS reconstructions of real-world driving logs. The reconstructed scenes are edited to stage each event, e.g., by inserting road users or hazards, scripting behaviors such as lead-vehicle braking or cut-ins, or setting traffic-light states. Following the official protocol, all policies are evaluated on all 280 scenarios with three seeds (mean ± std); baseline results are taken from the NavSafe-∞ paper.

PolicyDS↑SR↑DE↑Comf.↑TC↑VRUC↑TV↑TI↑
VLA-based
ReCogDrive-2B-RL53.64 ± 1.4230.95 ± 2.06173.96 ± 1.9085.10 ± 0.4935.33 ± 2.3118.33 ± 6.2932.42 ± 1.3927.78 ± 5.09
MTDrive-mtGRPO55.86 ± 0.5040.60 ± 0.90277.62 ± 3.1793.94 ± 0.4140.33 ± 2.8916.67 ± 1.4446.36 ± 0.9152.22 ± 3.85
AutoVLA51.05 ± 0.8029.52 ± 2.43186.02 ± 0.5865.37 ± 0.4733.00 ± 2.6515.00 ± 4.3329.70 ± 2.9236.67 ± 8.82
WAM-based
DriveVLA-W045.06 ± 0.6427.98 ± 1.80256.31 ± 1.0162.74 ± 1.1426.67 ± 2.317.50 ± 0.0033.94 ± 2.2937.78 ± 1.92
SimWAM57.53 ± 2.7237.50 ± 1.56160.54 ± 1.6497.19 ± 0.2741.33 ± 0.5826.67 ± 7.2240.61 ± 0.5227.78 ± 5.09
AffordDrive3D (Ours)58.56 ± 0.1845.12 ± 0.90166.99 ± 0.6790.87 ± 0.6244.33 ± 0.5823.33 ± 1.4454.55 ± 1.5742.22 ± 5.09

These edited, event-centered scenarios differ from the logged NAVSIM data and are considerably more challenging. AffordDrive3D achieves the highest driving score (DS) and success rate (SR) among all VLA- and WAM-based methods, demonstrating its generalization to safety-critical closed-loop driving.

Minor W2. Performance stability across seeds

We evaluate with 3 inference sampling seeds. NAVSIM v1 PDMS is 89.60 ± 0.03 after SFT and 91.33 ± 0.04 after RL, consistent with our reported results. This variance is far smaller than the improvements over competing methods, indicating that these gains are stable.

Minor W3. Inference latency and computational cost

All auxiliary branches, including the 1.3B video diffusion model and the geometry and affordance decoders, are discarded at inference. We measure the model inference on a single NVIDIA RTX 5090:

ComponentParamsLatency (ms)Compute
VLM (Qwen3-VL-2B, BF16)2.13B122.6 ± 0.2~13.7 TFLOPs
Action DiT (10 denoising steps, FP32)0.82B94.1 ± 0.4~0.11 TFLOPs
Others (queries, state encoder, etc.)0.08B––
Total3.03B217.9 ± 2.8~13.8 TFLOPs

Peak inference memory is 8.4 GiB (7.3 GiB for weights). Both latency and memory of the Action DiT, currently run in FP32 with sequential denoising, can be further reduced with lower precision or fewer sampling steps.

Minor W4. Long-horizon affordance prediction analysis

Accuracy decay at longer horizons is common to future prediction in driving. For example, 4D occupancy world models drop sharply from 1 s to 3 s on Occ3D-nuScenes [2] (OccWorld [3]: 25.78→10.51 mIoU; DOME [4]: 35.11→20.29 mIoU). In comparison, our decay is mild: collision-critical AP decreases from 0.918 to 0.732, and drivable-area AP stays high (0.986→0.971). The larger drop for collision-critical regions is expected, since they depend on the future behavior of dynamic agents, while drivable areas follow stable road structure. Despite this decay, future affordance supervision still benefits planning, improving TTC from 94.3 to 95.1 over current-frame supervision alone (see table in W1&Q1).

[1] Bao, Yuxin, et al. "NavSafe-$\infty $: Benchmarking Closed-Loop Driving Safety in Photorealistic Environments." arXiv preprint arXiv:2609.26618 (2026).

[2] Tian, Xiaoyu, et al. "Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving." Advances in neural information processing systems 36 (2023): 64318-64330.

[3] Zheng, Wenzhao, et al. "Occworld: Learning a 3d occupancy world model for autonomous driving." European conference on computer vision. Cham: Springer Nature Switzerland, 2024.

[4] Gu, Songen, et al. "Dome: Taming diffusion model into high-fidelity controllable occupancy world model." arXiv preprint arXiv:2410.10429 (2024).

Reviewer_SJ6oOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AffordDrive3D, a world-action model that jointly learns future RGB prediction, depth prediction, temporal driving affordances, and trajectory generation. The geometry branch predicts depth conditioned on future RGB latents, while the affordance branch predicts drivable areas and collision-critical regions. The model is trained with supervised fine-tuning followed by GRPO-based reinforcement learning.

Critical strengths
  1. The paper explicitly models temporal driving affordances, including drivable areas and collision-critical regions, alongside future appearance and geometry. This provides action-relevant supervision beyond generic scene prediction.

  2. The NAVSIM results are competitive, and the ablations suggest that affordance, RGB, and geometry supervision each contribute to planning performance.

Critical weaknesses
  1. Unclear novelty relative to DriveDreamer-Policy. The RGB and geometry components appear closely related to existing geometry-grounded world-action models. The paper should more clearly discuss the differences in future depth prediction, RGB-to-geometry conditioning, and the role of temporal affordance forecasting, supported by controlled comparisons where possible.

  2. Unclear consistency between predicted futures and RL-optimized actions. During RL, only the action diffusion model is updated, while the world model and auxiliary branches remain frozen. The paper does not evaluate whether the resulting trajectories remain consistent with the predicted drivable areas and collision-critical regions.

  3. Limited generalization and qualitative evaluation. All experiments are conducted on NAVSIM, and the qualitative results cover only a few scenarios. Zero-shot evaluation on nuScenes and visualizations of predicted affordance masks across diverse scenarios and future horizons would strengthen the evidence for generalization and prediction reliability.

Questions for authors
  1. How do the authors assess consistency between RL-optimized trajectories and the frozen future prediction branches? Can this consistency be quantified before and after RL?

  2. Can the authors provide zero-shot results on nuScenes, together with predicted drivable-area and collision-critical masks across multiple future horizons?

2 comments

We thank the reviewer for the comments.

W1. Novelty relative to DriveDreamer-Policy

Conditioning world-model branches on shared VLM tokens is a common design in VLM-based world-action models (e.g., DriveVLA-W0[1], DriveDreamer-Policy[2]). Beyond this, AffordDrive3D differs from DriveDreamer-Policy in two aspects:

  • Future geometry. DriveDreamer-Policy predicts only current depth with a separate pixel-space diffusion model. We forecast current and future depth conditioned on predicted RGB latents, keeping geometry consistent with future appearance, using lightweight single-pass 3D conv decoders.
  • Temporal affordance forecasting. Prior WAMs, including DriveDreamer-Policy, model what the future looks like but not which regions matter for ego planning. Our affordance supervision explicitly encodes where the ego vehicle can drive and which regions are collision-critical, aligning the representations with planning (largest single-branch gain, 88.1→89.2 PDMS, Table 3).

Q1&W2. Consistency between RL-optimized trajectories and predicted affordances

Since RL updates only the action DiT, all predicted futures (RGB, depth, and affordances) are identical before and after RL,consistency thus depends on how much RL alters the trajectory. We pair SFT and RL trajectories on all 12,146 navtest scenes, using the SFT model's own sampling variance as a reference:

RL vs. SFTSFT seed noise
ADE median / p900.15 / 0.28 m0.06 / 0.11 m
FDE (4 s) median / p900.19 / 0.43 m0.08 / 0.23 m
Scenes with ADE > 1 m0 (max 0.60 m)0
Endpoint lateral / longitudinal shift (median)0.03 / 0.17 m0.02 / 0.07 m
Lateral intent agreement97.4%99.0%
Longitudinal intent agreement94.3%97.8%

RL makes only small, local refinements: no trajectory deviates by more than 1 m, lateral shifts are at the noise level, and intent changes occur only between adjacent categories (e.g., straight vs. slight turn), never between left and right or accelerate and decelerate. Meanwhile, RL improves DAC (97.6→99.0) and TTC (95.2→96.3),indicating better compliance with drivable and collision constraints. Thus, RL refines trajectories within the future predicted by the frozen world model rather than departing from it.

Q2&W3. Zero-shot generalization and qualitative affordance visualization

Zero-shot transfer to nuScenes is confounded by substantial sensor differences. NAVSIM provides 8 surround-view cameras at 1920×1080, whereas nuScenes uses a 6-camera rig at 1600×900 (cropped from 1600×1200). The two datasets therefore differ in camera layout, intrinsics, and extrinsics, so zero-shot results would primarily reflect sensor-rig shift rather than planning generalization.

Instead, we evaluate generalization to closed-loop, safety-critical driving on NavSafe-∞ [3], a photorealistic benchmark of 280 event-centered scenarios spanning 28 safety-critical event types, built on 3DGS reconstructions of real-world driving logs. The reconstructed scenes are edited to stage each event, e.g., by inserting road users or hazards, scripting behaviors such as lead-vehicle braking or cut-ins, or setting traffic-light states. Following the official protocol, all policies are evaluated on all 280 scenarios with three seeds (mean ± std); baseline results are taken from the NavSafe-∞ paper.

PolicyDS↑SR↑DE↑Comf.↑TC↑VRUC↑TV↑TI↑
VLA-based
ReCogDrive-2B-RL53.64 ± 1.4230.95 ± 2.06173.96 ± 1.9085.10 ± 0.4935.33 ± 2.3118.33 ± 6.2932.42 ± 1.3927.78 ± 5.09
MTDrive-mtGRPO55.86 ± 0.5040.60 ± 0.90277.62 ± 3.1793.94 ± 0.4140.33 ± 2.8916.67 ± 1.4446.36 ± 0.9152.22 ± 3.85
AutoVLA51.05 ± 0.8029.52 ± 2.43186.02 ± 0.5865.37 ± 0.4733.00 ± 2.6515.00 ± 4.3329.70 ± 2.9236.67 ± 8.82
WAM-based
DriveVLA-W045.06 ± 0.6427.98 ± 1.80256.31 ± 1.0162.74 ± 1.1426.67 ± 2.317.50 ± 0.0033.94 ± 2.2937.78 ± 1.92
SimWAM57.53 ± 2.7237.50 ± 1.56160.54 ± 1.6497.19 ± 0.2741.33 ± 0.5826.67 ± 7.2240.61 ± 0.5227.78 ± 5.09
AffordDrive3D (Ours)58.56 ± 0.1845.12 ± 0.90166.99 ± 0.6790.87 ± 0.6244.33 ± 0.5823.33 ± 1.4454.55 ± 1.5742.22 ± 5.09

AffordDrive3D achieves the highest driving score (DS) and success rate (SR) among all VLA- and WAM-based methods, with the highest scores on TC and TV, which show safety-critical abilities. We will add more qualitative results, including predicted affordance masks across diverse scenarios and future horizons.

[1]Li, Yingyan, et al. "Drivevla-w0: World models amplify data scaling law in autonomous driving." International Conference on Learning Representations. Vol. 2026. 2026.

[2]Zhou, Yang, et al. "Drivedreamer-policy: A geometry-grounded world-action model for unified generation and planning." arXiv preprint arXiv:2604.01765 (2026).

[3]Bao, Yuxin, et al. "NavSafe-$\infty $: Benchmarking Closed-Loop Driving Safety in Photorealistic Environments." arXiv preprint arXiv:2609.26618 (2026).

Reviewer_yPkfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AffordDrive3D augments a vision-language driving policy with supervision for current and future drivable areas, collision-critical regions, future RGB, and depth. Affordance targets are constructed from maps, annotations, and trajectory simulations. Joint supervised training is followed by reinforcement learning of the action head; auxiliary predictions need not be decoded at inference. The paper reports 91.3 PDMS and 89.9 EPDMS on NAVSIM and ablations showing benefits from the auxiliary objectives.

Critical strengths
  1. Reasonable motivation: Adding supervision for driving-relevant regions to RGB and depth modeling is intuitively well motivated.
  2. Empirical improvements: Without RL, affordance supervision improves PDMS from 88.1 to 89.2, and the full model reaches 89.6.
Critical weaknesses
  1. Limited methodological novelty: The main addition is heuristic affordance supervision within an existing world-action modeling pipeline; the methodological advance beyond this combination is unclear.
  2. Insufficient label validation: No human evaluation is reported, and the limited visualizations do not establish that the heuristic drivable areas capture the claimed affordances; planning gains and AP against the same labels cannot resolve this concern.
Questions for authors
  1. Missing recent baselines: Accepted methods including SparseDriveV2 (ECCV 2026), DrivoR (CVPR 2026), and SGDrive (CVPR 2026) are omitted, weakening the SOTA claim.
  2. Unclear evaluation protocol: It is unclear whether the NAVSIM v2 results follow the standard two-stage protocol and use a consistent metric implementation across methods.
1 comment

We thank the reviewer for the comments.

W1. Methodological novelty

Existing WAMs model how the future looks, through RGB, latent features, or geometry of the entire future scene, but not which parts of the future matter for ego motion. Our core novelty is to make this action relevance an explicit world-modeling target:

  • Driving affordances as world-model outputs. To our knowledge, AffordDrive3D is the first WAM that densely forecasts drivable and collision-critical regions over current and future frames. These targets are grounded in the ego’s feasible maneuvers rather than in object categories, so the model learns which future regions constrain its actions, not merely what the scene contains.
  • RGB-conditioned future geometry. Instead of predicting geometry independently, we decode depth from the denoised latent of the RGB world model. This couples anticipated appearance with 3D structure.

W2. Affordance label validation

Our labels are deterministically computed from NAVSIM's ground-truth annotations, HD map, and official simulator, rather than from manual heuristics.

  • Collision-critical regions are obtained by rolling out candidate trajectories in the official NAVSIM v1 simulator (LQR controller + kinematic bicycle model, log-replayed agents) and marking objects involved in collisions.
  • Drivable areas are defined by reachability on NAVSIM's ground-truth roadblock graph (lawful lanes and intersection branches, excluding sidewalks and oncoming lanes).

We construct the labels this way so that they directly reflect what matters for ego planning, rather than generic scene semantics. The drivable-area target covers only road surface the ego vehicle can legally reach, not all visible road. For collision-critical regions, the candidate trajectory set approximates the space of feasible ego maneuvers, and rolling out all of them covers multiple plausible futures beyond the single logged one. As a result, only agents and objects that constrain the ego vehicle's possible actions are highlighted, rather than all objects in view.

The full affordance label construction pipeline is detailed in Appendix A.1. We will add more label visualizations across diverse scenarios and future horizons.

Q1. Missing recent baselines

We thank the reviewer and will include these methods in Table 1.

SGDrive is the most comparable method, sharing a 2B VLM backbone with RL fine-tuning. AffordDrive3D outperforms it on NAVSIM v1 (91.3 vs. 91.1 PDMS), with a notably higher DAC (99.0 vs. 97.8).

SparseDriveV2 and DrivoR follow a different paradigm: they score a dense set of trajectory candidates with learned PDM sub-score heads, whereas we generate trajectories with a world-action model. On safety and compliance metrics, our method achieves a higher DAC than both (99.0 vs. 98.4 / 98.3) and a clearly higher TTC than SparseDriveV2 (96.3 vs. 95.0), reflecting the benefit of forecasting drivable and collision-critical regions.

Q2. NAVSIM v2 evaluation protocol

Our NAVSIM v2 results report EPDMS on the navtest split, following the official NAVSIM v2 evaluation protocol and scoring metrics for EPDMS. We will include more details of the evaluation and metrics.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.