DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models
DeepThinkVLA is a Vision-Language-Action (VLA) model designed to enhance the reasoning capabilities of robotic agents through explicit deliberation (Chain-of-Thought). It refactors the policy into a hybrid-attention decoder that separates autoregressive reasoning from parallel action generation.
- Paper: DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models
- Repository: https://github.com/OpenBMB/DeepThinkVLA
Overview
DeepThinkVLA introduces two key innovations for effective Chain-of-Thought (CoT) in robotics:
- Decoding Alignment: A hybrid-attention decoder that pairs causal attention for language (reasoning tokens) with bidirectional attention for parallel action decoding.
- Causal Alignment: A two-stage SFT-then-RL pipeline that aligns the reasoning-action chain with sparse task-success rewards, ensuring that the generated "thoughts" are causally linked to task success.
The model achieves a 97.0% success rate on the LIBERO benchmark and demonstrates strong robustness under distribution shifts in LIBERO-Plus (79.0% zero-shot success).
Architecture
DeepThinkVLA inserts a <think> segment between visual observations and robot actions. Starting from the pi0-FAST checkpoint, it uses a 2.9B parameter hybrid decoder. Reasoning tokens are generated autoregressively, after which the model switches to parallel decoding for action chunks to maintain low latency.
Performance
- LIBERO: 97.0% average success rate (Object 99.0, Spatial 96.6, Goal 96.4, Long 96.2).
- LIBERO-Plus: 79.0% zero-shot robustness.
- RoboTwin 2.0: 59.3% success rate.
Setup and Evaluation
Please refer to the official GitHub repository for detailed instructions on environment setup, dataset acquisition, and running evaluation scripts.
Example evaluation command:
bash scripts/eval.sh --pretrained_checkpoint yinchenghust/deepthinkvla_libero_cot_sft
Citation
@article{yin2025deepthinkvla,
title={DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models},
author={Yin, Cheng and Lin, Yankai and Xu, Wang and Tam, Sikyuen and Zeng, Xiangrui and Liu, Zhiyuan and Yin, Zhouping},
journal={arXiv preprint arXiv:2511.15669},
year={2025}
}