Organizations: Bosch Corporate Research, China. · Institute for AI Industry Research (AIR), Tsinghua University, China. · Rensselaer Polytechnic Institute, USA.
End-to-end (E2E) driving has become a cornerstone of both industry deployment and academic research, offering a single learnable pipeline that maps multi-sensor inputs to actions while avoiding hand-engineered modules. However, the reliability of such pipelines strongly depends on how well they handle uncertainty: sensors are noisy, semantics can be ambiguous, and interaction with other road users is inherently stochastic. Uncertainty also appears in multiple forms: classification vs. localization, and, crucially, in both static map elements and dynamic agents. Existing E2E approaches model only static-map uncertainty, leaving planning vulnerable to overconfident and unreliable inputs. We present UniUncer, the first lightweight, unified uncertainty framework that jointly estimates and uses uncertainty for both static and dynamic scene elements inside an E2E planner. Concretely: (1) we convert deterministic heads to probabilistic Laplace regressors that output per-vertex location and scale for vectorized static and dynamic entities; (2) we introduce an uncertainty-fusion module that encodes these parameters and injects them into object/map queries to form uncertainty-aware queries; and (3) we design an uncertainty-aware gate that adaptively modulates reliance on historical inputs (ego status or temporal perception queries) based on current uncertainty levels. The design adds minimal overhead and drops throughput by only ∼0.5 FPS while remaining plug-and-play for common E2E backbones. On nuScenes (open-loop), UniUncer reduces average L2 trajectory error by 7%. On NavsimV2 (pseudo closed-loop), it improves overall EPDMS by 10.8% and notable stage two gains in challenging, interaction-heavy scenes. Ablations confirm that dynamic-agent uncertainty and the uncertainty-aware gate are both necessary.
Figures & tables
Fig. 1: Comparison between (a) a standard E2E autonomous driving model and (b) our uncertainty-aware E2E model. Our model introduces uncertainty estimation in both static and dynamic branches, and employs an uncertainty-aware gating mechanism to modulate historical information based on current uncertainty levels.
Fig. 2: Overview of the UniUncer framework. UniUncer extends the static and dynamic branches with a unified uncertainty estimation head to predict Laplace parameters. These parameters are encoded and fused with the original queries to form uncertainty-aware queries, which serve as input to the planner and gate historical information (e.g., ego state and temporal queries) via an uncertainty-aware gate.
Method
Reference
Input
Img. Backbone
L2 (m) ↓
Col. Rate (%) ↓
1s
2s
3s
Avg.
1s
2s
3s
Avg.
ST-P3 [ 13 ]
ECCV 2022
Camera
EffNet-b4
1.33
2.11
2.90
2.11
0.23
0.62
1.27
0.71
OccNet [ 39 ]
ICCV 2023
Camera
ResNet-50
0.45
0.70
1.04
0.73
0.62
0.58
0.63
0.61
UniAD [ 14 ]
CVPR 2023
Camera
ResNet-101
0.45
0.70
1.04
0.73
0.62
0.58
0.63
0.61
VAD [ 20 ]
ICCV 2023
Camera
ResNet-50
0.41
0.70
1.05
0.72
0.03
0.19
0.43
0.21
GenAD [ 48 ]
ECCV 2024
Camera
ResNet-50
0.36
0.83
1.55
0.91
0.06
0.23
1.00
0.43
TABLE I: Open-loop E2E performance benchmark on nuScenes.
Method
Backbone
Stage
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
HC ↑
EC ↑
EPDMS Stage ↑
EPDMS ↑
LTF [ 5 ]
ResNet-34
Stage 1
96.2
79.6
99.1
99.6
84.1
95.1
94.2
97.6
79.1
62.3
23.1
ResNet-34
Stage 2
77.8
70.4
84.3
98.1
85.1
75.7
45.4
95.7
76.0
37.5
DiffusionDrive [ 31 ]
ResNet-34
Stage 1
96.3
84.7
98.7
99.3
84.3
95.3
94.4
97.8
76.4
65.5
25.9
ResNet-34
Stage 2
79.5
71.2
84.9
98.3
86.5
75.6
46.9
96.2
70.9
39.6
DiffusionDrive + Ours
ResNet-34
Stage 1
96.8
82.9
98.8
99.3
83.6
95.8
96.0
97.6
77.3
64.1
28.7
ResNet-34
Stage 2
81.5
74.4
86.9
98.1
85.2
78.6
49.2
95.7
74.2
43.4
TABLE II: Pseudo Close Loop E2E Performance Benchmark on the navhard two-stage test from NavsimV2.
Method
D-Uncer
Uncer-Gate
Stage
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
HC ↑
EC ↑
EPDMS Stage ↑
EPDMS ↑
DiffusionDrive [ 31 ]
Stage 1
96.3
84.7
98.7
99.3
84.3
95.3
94.4
97.8
76.4
65.4
25.9
Stage 2
79.5
71.2
84.9
98.3
86.5
75.6
46.9
96.2
70.9
39.6
+D-Uncer
✓
Stage 1
95.9
85.6
98.7
99.3
84.5
95.1
94.9
97.8
76.0
65.5
26.7
Stage 2
79.2
72.3
85.9
98.2
86.6
76.2
48.2
96.3
70.1
40.6
+D-Uncer +Uncer-Gater
✓
✓
Stage 1
96.8
82.9
98.8
99.3
83.6
95.8
96.0
97.6
77.3
64.1
28.7
Stage 2
81.5
74.4
86.9
86.9
85.2
78.6
49.2
95.7
74.2
43.4
TABLE III: Ablation study on the effectiveness of dynamic uncertainty and the uncertainty-aware gate on the Navhard two-stage test.
Method
S-Uncer
D-Uncer
Uncer-Gate
L2 (m) ↓
Col. Rate (%) ↓
1s
2s
3s
Avg.
1s
2s
3s
Avg.
SparseDrive [ 36 ]
0.29
0.58
0.96
0.61
0.01
0.05
0.18
0.08
+S-Uncer
✓
0.30
0.58
0.95
0.61
0.01
0.05
0.18
0.08
+D-Uncer
✓
0.27
0.53
0.90
0.57
0.02
0.11
0.21
0.11
+S-Uncer +D-Uncer
✓
✓
0.27
0.53
0.88
0.56
0.01
0.06
0.19
0.09
+S-Uncer +D-Uncer +Uncer-Gate
✓
✓
✓
0.27
0.53
0.90
0.57
0.00
0.04
0.17
0.07
TABLE IV: Ablation study on the effectiveness of static uncertainty, dynamic uncertainty, and the uncertainty-aware gate on nuScenes.
Fig. 3: Visualization of the Uncertainty-aware Gate (fine-grained feature- time matrix mode) for Ego Status. The upper row illustrates a complex scenario, while the lower row corresponds to a simple scenario. Driving commands include L (left turn), S (keep straight), R (right turn), and NA (not available). Vx and Vy denote velocities along the x - and y -axes, respectively, while ax and ay represent accelerations along the x - and y -axes.
Method
Number of Params (M)
FPS
SparseDrive [ 36 ]
86.2
6.6
SparseDrive + Ours
91.8
6.1
TABLE V: Efficiency of Our Approach. All methods are tested on a single RTX 3090 GPU.
Fig. 4: Qualitative results on the nuScenes dataset. With unified uncertainty estimation, the predicted trajectory from our model is closer to the ground-truth trajectory compared to the baseline model. Ellipses on map vertices and dynamic objects represent the scale parameter b along the x - and y -axes. For dynamic objects, the uncertainty is lower on the visible side of the vehicle compared to the invisible side. Best viewed in the digital version.
Fig. 5: Qualitative results on the Navhard two-stage test. Uncertainty estimation enables the planner to generate trajectories that are closer to the ground-truth (GT) trajectory in general. It demonstrates that uncertainty estimation for dynamic agents is critical for collision avoidance in challenging pseudo closed-loop scenarios rendered by 3DGS, which can introduce synthetic artifacts (e.g., the bus and vehicle in the second row). By assigning higher uncertainty to such potentially unreliable detections, the planner adopts a more conservative strategy, successfully avoiding over-reaction and improving robustness. Better view in digital version.
Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.
End-to-end planning has shown strong promise for autonomous driving, but most existing methods are designed for structured urban roads and generalize poorly to unstructured mining environments. In such settings, weak road structure, terrain-induced occlusions, degraded visibility, and large unobserved regions make safe planning particularly challenging. To address these challenges, we propose UnsDrive, an end-to-end planner designed for unstructured mining scenes. UnsDrive builds an unknown-aware occupancy representation that explicitly models occupied, free, and unknown space using multi-frame visibility cues, and conditions a flow-matching planner on this representation to generate multimodal future trajectories. To improve safety under partial observability, we further introduce an occupancy trajectory consistency loss and an uncertainty-aware trajectory scorer that penalize trajectories entering non-traversable or unobserved regions. We also present MineLoop, a mining-oriented closed-loop simulator for evaluating autonomous driving under irregular road geometry, degraded visibility, heavy-vehicle interactions, and mining-specific operational constraints. Experiments in both open-loop and closed-loop settings show that UnsDrive consistently outperforms strong baselines in trajectory accuracy, collision avoidance, and long-horizon driving robustness. These results demonstrate the value of explicit unknown-space reasoning for autonomous driving in unstructured mining environments.
Nanxin Zeng, Ruiqi Song, Xiangyu Guo +2
Institute of Automation, Chinese Academy of Sciences Beijing, China · Waytous Inc. Beijing, China
Learning robust driving policies that generalize beyond their training domain remains a central challenge in motion planning. We propose an uncertainty-driven self-play scheme, in which half of the agents are rewarded for increasing their neighbors' epistemic uncertainty, while the other half learn to drive safely as the exploring agents continuously seek out their weaknesses. Epistemic uncertainty indicates where the policy is least confident and serves as an automatic proxy for scenario difficulty. Exploring agents use this signal to seek out hard cases, generating a challenging curriculum without manual scenario design. The resulting policy outperforms standard self-play, achieving best-in-class zero-shot transfer to nuplan Val14 (99.12% vs. 97.97%) and Argoverse 2 (82.39% vs. 78.03%) without fine-tuning, while also improving performance over standard self-play on Waymo. Our diversity-driven approach achieves 85% higher LSTM hidden-state coverage than the baseline, indicating a much richer set of behaviors. Further analysis shows that epistemic uncertainty serves as a calibrated difficulty detector, with highest-uncertain agents exhibiting 14.8 times the combined failure rate across collision, off-road, and DNF, of lowest-uncertain agents, enabling automatic prioritization of challenging scenarios without ground-truth labels.
The end-to-end (E2E) paradigm, which maps sensor inputs directly to driving decisions, has recently attracted significant attention due to its unified modeling capability and scalability. However, ensuring safety in this unified framework remains one of the most critical challenges. In this work, we propose SafeDrive, an E2E planning framework designed to perform explicit and interpretable safety reasoning through a trajectory-conditioned Sparse World Model. SafeDrive comprises two complementary networks: the Sparse World Network (SWNet) and the Fine-grained Reasoning Network (FRNet). SWNet constructs trajectory-conditioned sparse worlds that simulate the future behaviors of critical dynamic agents and road entities, providing interaction-centric representations for downstream reasoning. FRNet then evaluates agent-specific collision risks and temporal adherence to drivable regions, enabling precise identification of safety-critical events across future timesteps. SafeDrive achieves state-of-the-art performance on both open-loop and closed-loop benchmarks. On NAVSIM, it records a PDMS of 91.6 and an EPDMS of 87.5, with only 61 collisions out of 12,146 scenarios (0.5%). On Bench2Drive, SafeDrive attains a 66.8% driving score.