yuki-hahaha/AUGUstus

Paper · https://arxiv.org/pdf/2610.16985 · Project Page · https://augustus-ad.github.io/

Figure 1. The difference between AUGUstus and other autonomous-driving VLA paradigms.

Figure 2. Overview of AUGUstus.

Model checkpoints

AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding.

This Hugging Face repository contains the SFT, GRPO stage-I, and GRPO stage-II checkpoints. Source code and the project page are distributed in separate GitHub repositories.

Checkpoint directory	Stage	ODE denoising steps	NAVSIM v1 PDMS
SFT/	Supervised fine-tuning	1	90.64
RL/stage1_94.0/	GRPO stage I	1	94.03
RL/stage2_94.6/	GRPO stage II	3	94.64

NAVSIM v1 evaluation uses 12,146 samples. Use RL/stage2_94.6/ with 3 denoising steps for the 94.6-PDMS configuration.

Each checkpoint includes model.safetensors, Hugging Face backbone configuration, tokenizer, processor, chat template, and generation settings. config.yaml contains the model architecture, inference settings, data paths, and normalization settings used by the current AUGUstus code. action_norm_stats.json provides action normalization statistics; the checkpoint's normalization buffers are restored when its weights are loaded. Each checkpoint has exactly one YAML configuration. SFT and GRPO training recipes are maintained in the code repository.

Use with the source code

Install the separately distributed AUGUstus source and dependencies, acquire the Qwen3.5-2B backbone, and prepare the NAVSIM test and metric caches as described in the code repository. The code repository has adjacent AUGUstus_SFT/ and AUGUstus_GRPO/ directories at its root.

From the SFT directory:

cd /path/to/AUGUstus-code/AUGUstus_SFT
source env.sh
export CHECKPOINT=/path/to/downloaded-AUGUstus-model/RL/stage2_94.6
CONFIG="$CHECKPOINT/config.yaml" \
GPUS=0 NUM_INFERENCE_STEPS=3 GENERATE_COT=0 \
  bash scripts/test_checkpoint.sh

CHECKPOINT must be exported: the architecture YAML uses it to locate the bundled normalization JSON. Set QWEN35_2B_PATH, AUGUSTUS_DATA_ROOT, and AUGUSTUS_EXP_ROOT before sourcing env.sh when your data and backbone use different storage locations. Evaluation output is written beneath the selected checkpoint unless OUTPUT_DIR is overridden.

These files contain the complete custom AUGUstus planner state, including action and time modules. Use the AUGUstus loader; direct loading as a plain Qwen AutoModel does not assemble the planner. The configurations use the current use_action_first and action_first_generate_cot keys. SFT and stage I default to one denoising step; stage II defaults to three.

Terms and citation

The manuscript states: "This work is solely for non-commercial academic research." Third-party backbone terms remain applicable. The source-code repository's license is separate from model-weight terms.

@misc{ye2026augustusautonomousdrivingvla,
  title={AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding},
  author={Yuqi Ye and Caigao Jiang and Zhuonan Wang and Congyuan Pei and Chao Qu and Changhao Peng and Junhong Lin and Qize Yu and Zetian Song and Guangzhao Fu and Chongruo Wu and Shiyu Huang and Shuchang Zhou and Honggou Yang and Wei Gao},
  year={2026},
  eprint={2610.16985},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.16985},
}

Contact: yeyuqi0303@stu.pku.edu.cn.