AutoDrive-OPD-E057V-progress-teacher

Progress expert Teacher. VGS-GRPO from the SFT model E055S on the same 12K navtrain scenes and settings as E057D.

Part of an independent reproduction and extension of Driving-OPD on NAVSIM. Full-parameter Qwen3-VL-4B-Instruct (revision ebb281ec70b05090aa6165b016eac8ec08e71b17).

navtest PDMS navtest EPDMS (guarded) navhard EPDMS
0.9105 0.8978 -

Scores are fixed-endpoint, greedy single-trajectory results on all 12,146 navtest scenes (0 parse failures).

Input and output

  • Input: the current front camera image (CAM_F0) and a text prompt with the navigation command, velocity, longitudinal acceleration, yaw rate and four past ego poses (navsim/agents/driving_opd/representation.py::make_prompt).
  • Output: one <trajectory> block with 8 [x, y, yaw] points at 0.5 s intervals in the current ego frame.
  • Evaluation settings: greedy decoding, max_new_tokens=512, min_pixels=262144, max_pixels=1048576.

Code

github.com/aba122/AutoDrive-OPD (private), branch driving-opd-release-20261119, commit 671e7ce. Generation: scripts/evaluation/generate_driving_opd.py; scoring: scripts/evaluation/score_driving_opd.py; training: navsim/planning/training/driving_opd_sft.py and navsim/planning/training/driving_opd_posttrain.py.

Data and license

Base model Qwen3-VL-4B-Instruct (Apache-2.0). Trained on NAVSIM / OpenScene navtrain scenes derived from nuPlan; use is subject to the nuPlan dataset terms.

File hashes (SHA-256)

file sha256
added_tokens.json c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680
config.json 8208922cd99b96b08a1f4b44b73a37661d849a1fd58c564e1f48c7a94da44a0e
generation_config.json 3ff2e83a0510cccccc85c8c96c7df4207985c098b526d798bc2bc68e50bb1a41
model-00001-of-00003.safetensors 4208ca7d3fb2357d8645eaed5b81d8cf864e97376498682f90b5e00f1a0bcc61
model-00002-of-00003.safetensors 98720b6181344e9c0382131d084c00802b34153e1eabe7c9ea3742bc4db0fa81
model-00003-of-00003.safetensors c1fd229e45a4cc014a5050c8da3bcb297fce4f99b5fa0a3dbb71d9ab8acdc973
model.safetensors.index.json 3b73100a43c7a253625a355b669ea12998739a9ead79600acab76558d6b1aada
preprocessor_config.json 70b360c2f47063b3dc1b73eb94cace5538947b565545ef44781e5643349d77c5
special_tokens_map.json 76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
tokenizer.json aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
tokenizer_config.json 76aecc1df638ee1adf284cfcbf05d1d24905d44e576ffb37876b592908b13977
video_preprocessor_config.json 59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1
vocab.json ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
Downloads last month
-
Safetensors
Model size
4B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for wulin222/AutoDrive-OPD-E057V-progress-teacher

Finetuned
(571)
this model