AutoVLA-D: Learn a Trajectory-Geometric Condition from Reasoning for Autonomous Driving VLA
Authors: Yuguang Yang, Zhewen Tan, Canyu Chen, Linlin Yang, Cheng Chi, Chunyang Liu, Kehua Sheng, Jinyu Yang, Baochang Zhan, Bo Zhang, Yan Wang, Xianbin Cao
Organizations: School of Electronic and Information Engineering, Beihang University · Institute for AI Industry Research (AIR), Tsinghua University · School of Computer Science and Engineering, Beihang University · National College for Excellent Engineers, Beihang University · State Key Laboratory of Media Convergence and Communication, Communication University of China · School of Information, Renmin University of China · Didi · Harbin Institute of Technology, Shenzhen · Hangzhou Innovation Research Institute, Beihang University
Autoregressive VLA planning usually predicts continuous trajectories as the final output, but driving instructions often describe actions rather than ego states, which weakens direct conditioning on the final continuous trajectory. We address this gap with TrajCond, which introduces a discrete trajectory-geometric condition for VLA planning. Specifically, we represent TrajCond as a 1D-Brohan differential action trajectory and learn it with a two-stage training framework before final continuous trajectory prediction. With the simultaneous autoregressive generation of discrete and numeric trajectories, we name our method AutoVLA-D (Double). Experiments on NAVSIM navtest, nuScenes, and NeuroNCAP show that discrete trajectories are weaker as direct outputs but more effective as TrajCond, consistently improving final continuous trajectory prediction across closed-loop and open-loop driving benchmarks. Code is available at https://github.com/ShivaTeam/AutoVLA-D.
Figures & tables
Figure 1: Pipeline of AutoVLA-D.
Method
Base Model
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
ImagiDrive-S
InternVL2.5-4B
98.6
96.2
80.5
94.5
100.0
87.4
ReCogDrive
InternVL3-8B
98.2
97.8
83.5
95.2
100.0
89.6
ReCogDrive †
Qwen3-VL-4B
98.2
97.7
86.0
97.7
100.0
90.7
DriveVLA-W0
Emu3-8B
98.7
99.1
83.3
95.3
99.3
90.2
AutoVLA
Qwen2.5-VL-3B
98.4
95.6
81.9
98.0
99.9
89.1
AdaThinkDrive
InternVL3-8B
98.4
97.8
84.4
95.2
100.0
90.3
Table 1: NAVSIM navtest PDMS results under the public metric setting. † denotes results reproduced by us. ∗ denotes concurrent work.
Method
Base Model
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
C ↑
EC ↑
EPDMS ↑
ReCogDrive
InternVL3-8B
98.3
95.2
99.5
99.8
87.1
97.5
96.6
98.3
86.5
83.6
DiffusionDrive
–
98.2
95.9
99.4
99.8
87.5
97.3
96.8
98.3
87.7
84.5
Curious-VLA
Qwen2.5-VL-3B
98.4
96.9
99.2
99.8
88.5
97.9
96.9
98.1
81.5
85.3
Curious-VLA †
Qwen3-VL-4B
98.5
96.7
99.37
99.9
87.3
98.0
96.5
98.3
83.0
84.9
DriveVLA-W0
Emu3-8B
98.5
99.1
98.0
99.7
86.4
98.1
93.2
97.9
58.9
86.1
FLARE ∗
Qwen3-VL-4B
98.8
96.6
99.5
99.9
87.9
98.2
95.8
98.4
87.5
86.3
Table 2: NAVSIM navtest EPDMS results following the official EPDMS protocol. AutoVLA-D uses the same 1-camera setting as the PDMS experiment and reports the completed EPDMS evaluation. † denotes results reproduced by us. ∗ denotes concurrent work. EPDMS denotes the bug-fixed EPDMS metric released by the official repository of NAVSIM v2.
Table 4
Table 5: Progressive performance improvement on NAVSIM with Qwen3-VL-4B. We abbreviate Reason, Continuous, Brohan, and epochs as R, C, B, and ep., respectively. The right panel reports the corresponding EPDMS/PDMS scores.
Figure 2: Collision examples and failed NC×DAC product rates on NAVSIM navtest. The left three panels show representative collision-related scenes for the three high-level intents, and the right panel reports the failed NC×DAC product rate for Continuous and Brohan TrajCond. Full NC/DAC/Product counts are provided in Supplement Table 14 .
Figure 3: TrajCond form analysis for action tokens. Each codebook-size group pairs the reconstruction-error curve with the corresponding EPDMS/PDMS bars. The left axis reports trajectory reconstruction error; the right axis reports EPDMS/PDMS. Due to space limits, we move codebook-size=384 to Supplement Fig. 8 .
Figure 4: t-SNE visualization of hidden states from the discrete trajectory used as TrajCond. Left: points are colored by instruction motion direction. Right: points are colored by speed behavior.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Illustration of local-frame differential motion. The displacement between τt−1 and τt is first observed in the t=0 ego frame, then rotated by R(−θt−1) into the local frame at step t−1 to obtain (Δxt,Δyt) .
Figure 7: Trajectory tokenization illustration Copied From Trajeglish Philion et al. [2024] . An example trajectory is shown in green. The raw representation of the tokenized trajectory is shown as boxes with blue outlines. States that have yet to be tokenized are light green. Token templates are optimized to minimize the error between the tokenized trajectories and the raw trajectories.
Method
Base Model
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
Ego-MLP
–
93.1
78.3
63.2
84.0
100.0
66.4
UniAD
–
97.7
92.8
79.2
92.8
100.0
84.0
DiffusionDrive
–
98.2
96.2
82.2
94.7
100.0
88.1
WoTE
–
98.5
96.8
81.9
94.4
99.9
88.3
Qwen2.5-VL
Qwen2.5-VL-7B
97.8
92.1
78.3
92.8
100.0
83.3
InternVL3
InternVL3-8B
97.0
92.4
78.9
91.8
100.0
83.3
Appendix
Table 7: Additional baselines on NAVSIM navtest under PDMS. These rows are moved from the main table to keep the main comparison focused on stronger VLA methods.
Method
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
C ↑
EC ↑
EPDMS ↑
Ego-MLP
93.1
77.9
92.7
99.6
86.0
91.5
89.4
98.3
85.4
64.0
VADv2
97.3
91.7
98.2
99.9
77.6
92.7
66.0
100.0
97.4
76.6
TransFuser
96.9
89.9
97.8
99.7
87.1
95.4
92.7
98.3
87.2
76.7
HydraMDP++
97.2
97.5
99.4
99.6
83.1
96.5
94.4
98.2
70.9
81.4
DriveSuprim
97.5
96.5
99.4
99.6
88.4
96.6
95.5
98.3
77.0
83.1
ARTEMIS
98.3
95.1
98.6
99.8
81.5
97.4
96.5
98.3
–
83.1
Appendix
Table 8: Additional E2E baselines on NAVSIM navtest under EPDMS. These rows are moved from the main table to keep the main comparison focused on VLA methods.
Method
Stage 1
Stage 2
RFT
EPDMS ↑
PDMS ↑
Brohan (B)
R → B (6 ep.)
–
–
82.4
86.8
Continuous (C)
R → C (6 ep.)
–
–
83.9
88.5
Brohan TrajCond (1)
–
B → R → C (6 ep.)
–
84.0
89.2
Brohan TrajCond
R → B (3 ep.)
B → R → C (3 ep.)
–
85.6
90.8
Continuous + RFT
R → C (6 ep.)
–
✓
84.9
89.7
Brohan TrajCond + RFT
R → B (3 ep.)
B → R → C (3 ep.)
✓
86.8
92.0
Appendix
Table 9: Detailed progressive-analysis data behind Table 5 .
Figure 8: Complete trajectory-representation analysis across codebook sizes 384, 768, and 1536. The figure reports reconstruction error for each discretization method together with the corresponding PDMS/EPDMS performance bars.
Table 13: Absolute perplexity values for the pause-token replacement analysis. Lower is better.
Intent
NC zero count
DAC zero count
NC×DAC zero
Ours
C
Reduction
Ours
C
Reduction
Ours
C
Ours rate
C rate
Reduction
Go straight
160
263
39.16%
153
209
26.79%
308
464
2.54%
3.82%
33.62%
Turn left
29
52
44.23%
151
200
24.50%
178
246
1.47%
2.03%
27.64%
Turn right
35
53
33.96%
85
162
47.53%
118
204
0.97%
1.68%
42.16%
Appendix
Table 14: Intent-grouped NC/DAC zero-count comparison against the Continuous (C) baseline. Product rates use 12,146 samples as the denominator, and reduction is computed as (Baseline−Ours)/Baseline .
Figure 9: Additional collision-related examples for the NC/DAC analysis. The legend maps pink circular markers and square bars to Continuous, and green circular markers and square bars to Brohan TrajCond. The six panels preserve the original sample-image resolution after white-border cropping and show representative failure modes indicated by their panel labels: road-side collision, lane departure, excessive speed, and wrong direction.
School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China · State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University, Beijing 100084, China · School of Computer Science and Engineering, Southeast University, Nanjing 210096, China +2
School of Electronic Information Engineering, Beihang University · Institute for AI Industry Research (AIR), Tsinghua University · National College for Excellent Engineers, Beihang University +5