Reading list

Start here

1. Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail, NVIDIA, 2025. arXiv 2511.00088. Look for: how the Chain of Causation labels are built, the reward terms in the RL stage, and the open-loop versus closed-loop results.
2. Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning, NVIDIA, 2025. arXiv 2503.15558. The family of reasoning VLMs the backbone comes from. Look for: the four-stage training and how physical common sense is benchmarked.

Earlier driving models built on language models

3. EMMA: End-to-End Multimodal Model for Autonomous Driving, Waymo, 2024. arXiv 2410.23262. A large multimodal model that handles detection and motion planning in one model. Look for: everything expressed as text.
4. OpenDriveVLA: Towards End-to-end Autonomous Driving with Large Vision Language Action Model, 2025. arXiv 2503.23463. Evaluated on nuScenes. Look for: how 2D image features are aligned with 3D scene structure.

A map of the field

5. A Survey on Vision-Language-Action Models for Autonomous Driving, 2025. arXiv 2506.24044, with a companion list at Awesome-VLA4AD. Use it to place any new paper in the landscape rather than to read cover to cover.

How to read a paper in this list

Read in three passes, and stop when a pass answers your question.

1. Ten minutes. Abstract, the main figure, the results table and the conclusion. Write one sentence: what do the authors claim, and on what evidence?
2. One hour. Method and experiments, skipping proofs. For each table ask: open loop or closed loop, how many scenarios or episodes, and are differences larger than the stated error bars?
3. Reproduce a piece. Pick one equation or one block and implement it, or find it in the tutorial's reference code. A paper you cannot reduce to a small program is one you have not yet understood.

Keep a running table of every paper: input window, output, label source, headline metric, what was released. Comparing columns across papers is where the field's structure shows up.

A four-week path through the list

Code and weights

• Alpamayo repository: github.com/NVlabs/alpamayo. Its README says the repository is in limited maintenance and points to a newer release.
• Weights: nvidia/Alpamayo-R1-10B on Hugging Face. Access is gated.

Background and the wider literature

The tutorial lists the foundation papers (transformers, ViT, CLIP, VAE, diffusion, flow matching, ACT, Diffusion Policy) at the end of each module, and Driving VLA literature and Driving data and labels review the driving line of work (UniAD, GPT-Driver, DriveLM, DriveVLM, EMMA, OpenDriveVLA, DiffusionDrive, ReCogDrive). Manipulation VLAs (RT-2, OpenVLA, pi0, SmolVLA) and action tokenisation (FAST) are covered in VLA models; PPO, GRPO and DeepSeek-R1 in Reinforcement learning; chain of thought, self-consistency and ReAct in Reasoning; world models in World models; tokenizers (VQ-VAE, FSQ) in Discrete tokenizers; DDIM, guidance and diffusion policy in Sampling and guidance; sim-to-real in RL for real robots; DPO and offline RL in Preference learning and offline RL; LoRA, quantization and serving in Efficient fine-tuning; the onboard latency budget, token pruning and asynchronous execution in Deploying a driving VLA; conformal prediction and safety in Uncertainty and runtime safety; vehicle models, IDM and trajectory vocabularies in Planning and vehicle models; metrics and benchmarks in Evaluating driving models; cameras, ego-motion and DETR in Camera geometry and object queries.