Autonomous driving requires safe planning, but most learning-based planners lack explicit self-correction ability: once an unsafe action is proposed, there is no mechanism to correct it. Thus, we propose CorrectionPlanner, an autoregressive planner with self-correction that models planning as motion-token generation within a propose, evaluate, and correct loop. At each planning step, the policy proposes an action, namely a motion token, and a learned collision critic predicts whether it will induce a collision within a short horizon. If the critic predicts a collision, we retain the sequence of historical unsafe motion tokens as a self-correction trace, generate the next motion token conditioned on it, and repeat this process until a safe motion token is proposed or the safety criterion is met. This self-correction trace, consisting of all unsafe motion tokens, represents the planner's correction process in motion-token space, analogous to a reasoning trace in language models. We train the planner with imitation learning followed by model-based reinforcement learning using rollouts from a pretrained world model that realistically models agents' reactive behaviors. Closed-loop evaluations show that CorrectionPlanner reduces collision rate by over 20% on Waymax and achieves state-of-the-art planning scores on nuPlan.
Figures & tables
Figure 1: (a) Vanilla autoregressive models, no correction mechanism. (b) autoregressive models that generate tokens sequentially, with reasoning through language. (c) autoregressive models that generate subsequent tokens sequentially with self-correction through motion tokens.
Figure 2: Blue box is ego; red cross is collision; orange trajectory is the expert trajectory. Left: autoregressive without self-correction. Right: autoregressive with self-correction. Collision is avoided with self-correction before the collision by slowing down to yield to the right-turn vehicle and then re-accelerating.
Figure 3: (a) Policy architecture. A frozen reactive world model predicts future motions for the ego and other agents. Conditioned on scene history and map/navigation context, the planner is an autoregressive policy that generates the next ego motion token. When self-correction is triggered, previously rejected proposals are retained as a correction trace and integrated via a self-attention encoder to condition subsequent proposals, providing an explicit correction signal during both imitation learning and RL. (b) Self-correction. At each planning timestep, the policy proposes an ego motion token, and a learned collision critic evaluates whether it will cause a collision within a short horizon. If unsafe, the policy iteratively generates revised motion tokens conditioned on the correction trace until a safe token is found (or a maximum number of correction steps is reached). The resulting rollout contains both executed ego tokens and intermediate correction tokens, and the policy is further optimized with REINFORCE using a rule-based reward and KL regularization.
Reactive
Non-Reactive
Method
Collision ↓
Offroad ↓
Progression ↑
Collision ↓
Offroad ↓
Progression ↑
EasyChauffeur-PPO∗ ( Xiao et al., 2024 )
4.71
3.81
97.93
4.53
2.14
96.42
PlanT∗ ( Renz et al., 2022 )
2.94
1.65
95.85
3.02
1.43
95.48
LatentDrive∗ ( Xiao et al., 2025 )
3.27
2.33
98.70
3.11
1.94
99.21
SMART∗ ( Wu et al., 2024 )
2.36
0.87
91.33
4.30
0.86
90.87
Ours
1.68
0.94
94.23
2.43
0.92
95.75
Table 1: Comparison with baselines on Waymax simulator. Our method reduces the collision rate by over 20% . ∗ means we re-implement the method. We use bold and underline to denote the best and second best.
Val14
Test-Hard
Test-Random
Learning-Based Method
NR
R
NR
R
NR
R
PDM-Open ( Dauner et al., 2023 )
53.53
54.24
33.51
35.83
52.81
57.23
UrbanDriver ( Scheel et al., 2022 )
68.57
64.11
50.40
49.95
51.83
67.15
PlanTF ( Cheng et al., 2024b )
84.72
76.95
69.70
61.61
85.62
79.58
PLUTO ( Cheng et al., 2024a )
88.89
78.11
70.03
59.74
89.90
78.62
DiffusionPlanner ( Zheng et al., 2025 )
89.87
82.80
75.99
69.22
89.19
82.93
Table 2: Comparison with baselines on nuPlan under non-reactive (NR) and reactive (R) settings. Our method outperforms baselines. ∗ means we re-implement the method. We use bold and underline to denote the best and second best.
Type
Planner Score ↑
Collisions ↓
TTC ↑
Drivable ↑
Comfort ↑
Progress ↑
Speed ↑
SMART
90.03
2.70
91.79
99.02
99.34
91.68
97.38
Ours
91.22
2.04
92.25
98.76
99.31
91.49
97.49
Table 3: Planning performance comparison on Val14 NR.
Figure 4: Visualization on self-correction. (a),(c) and (e): our method w/o self-correction in inference. (b), (d) and (f): our method w/ self-correction. Orange trajectory is the log trajectory. Our method avoids the collision in these scenarios.
Method
Collision ↓
Progression ↑
Waymo
IL
2.34
95.77
IL (w/ corr.)
2.31
95.41
RL (w/o corr.)
2.21
94.98
CorrectionPlanner (w/o corr.)
2.20
95.02
CorrectionPlanner
1.68
94.23
Table 4: Comparison across training frameworks.
Figure 5: Effect of collision classification threshold and correction length on safety and progression.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Collision ↓
Progression ↑
Alternate Self-Correction
2.01
0.9501
Ours
1.68
0.9423
Appendix
Table 7: Comparison with alternate self-correction baseline on WOMD.
Method
Collision ↓
Offroad ↓
Progression ↑
EasyChauffeur-PPO ( Xiao et al., 2024 )
4.71
3.81
97.93
PlanT ( Renz et al., 2022 )
2.94
1.65
95.85
LatentDrive ( Xiao et al., 2025 )
3.27
2.33
98.70
SMART ( Wu et al., 2024 )
2.36
0.87
91.33
Zero Shot (w/o corr.)
2.93
1.40
93.64
Zero Shot (w/ corr.)
2.62
1.26
92.83
Appendix
Table 8: Zero-shot generalization ability of our method. Policy trained on the nuPlan dataset, evaluated on the Waymax simulator. The zero-shot (w/ corr.) is our method, and it achieves the second-lowest collision rate, alongside our method trained on WOMD, demonstrating zero-shot generalization and self-correction.
Method / Threshold
0.00
0.70
0.75
0.80
SMART
0.329
–
–
–
SFT
0.357
–
–
–
CorrectionPlanner
Correction length = 1
0.917
0.497
0.392
0.387
Correction length = 2
0.994
0.525
0.413
0.407
Correction length = 5
1.162
1.036
0.434
0.431
Appendix
Table 9: Waymo latency (seconds) under different collision classification thresholds per trajectory. The best-performing hyperparameters, with 0.75 as the threshold and 5 as the correction length, don’t significantly increase inference time. Experiment conducted on one NVIDIA H800 GPU.
k
Collision ↓
Progression ↑
1
2.19
96.53
3
1.97
94.99
5
1.68
94.23
7
1.63
91.42
Appendix
Table 10: Experiment results on different k when training the collision critic on WOMD.
Figure 6: Slow down and yield to the right-turning vehicle.
Figure 7: Slow down and yield to the right-turning vehicle, re-accelerate afterwards.
Figure 8: Adjust the speed in the left turn to avoid a collision.
Figure 9: Adjust the left turn to avoid a collision in the intersection.
Figure 10: Adjust the speed and yield to multiple left turn vehicle.