Collision Avoidance on CARLA Town02 (unseen)
100SR (Straight)InDRiVE
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| InDRiVEmode=Few-shot adaptation, extrinsic_reward_steps=10k, fine-tuning=task policy/value head only, frozen_components=pretrained world model and intrinsic module2025.12 | 100 | 0 | 100 | 0 | 93.57 | 6.43 | 85.55 | 14.45 | |
| RNDmode=Few-shot adaptation, extrinsic_reward_steps=10k, fine-tuning=task policy/value head only, frozen_components=pretrained world model and intrinsic module2025.12 | 95.2 | 4.8 | 91.23 | 8.77 | 88.45 | 11.55 | 80.29 | 19.71 | |
| DreamerV3mode=Few-shot adaptation, extrinsic_reward_steps=10k, fine-tuning=task policy/value head only, frozen_components=pretrained world model and intrinsic module2025.12 | 92.71 | 7.29 | 97.29 | 2.71 | 85.52 | 14.48 | 78.02 | 21.98 | |
| ICMmode=Few-shot adaptation, extrinsic_reward_steps=10k, fine-tuning=task policy/value head only, frozen_components=pretrained world model and intrinsic module2025.12 | 89.47 | 10.53 | 83.07 | 16.93 | 89.36 | 10.64 | 78.21 | 21.79 |