Source-linked AI summary

CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models

Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, Ankur Handa, Ming-Yu Liu, Donglai Xiang, Gordon Wetzstein, Tsung-Yi Lin

arXiv:2503.22020v1cs.CVcs.AIcs.LGcs.RO

TL;DR

VLAs generally lack explicit intermediate reasoning for complex manipulation because they map observations directly to actions. CoT-VLA generates future subgoal images as visual chain-of-thought steps before producing short action sequences, and experiments report improved performance across simulation and real-world tasks. The approach also enables action-less videos to train visual reasoning, but intermediate image generation creates substantial inference overhead.

  • Problem

    Existing VLAs directly map observations to actions without explicit intermediate reasoning steps for complex manipulation tasks.

  • Method

    CoT-VLA autoregressively generates a future subgoal image as an intermediate visual reasoning step, then conditions a short action sequence on the observation and generated goal.

  • Results

    CoT-VLA improves policy performance over prior VLA approaches in simulation and real-world experiments, with state-of-the-art performance across multiple robot platforms and tasks.

  • Takeaways & Limitations

    Using subgoal images lets CoT-VLA leverage information already present in robot manipulation data and potentially use abundant action-less video for visual reasoning.

  • Takeaways & Limitations

    Generating 256 intermediate image tokens before action tokens causes a 7× average slowdown with an action chunk size of 10.

Abstract

from arXiv · show

Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input--output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks. Project website: https://cot-vla.github.io/

1. Introduction

Existing VLAs map observations and language directly to actions, without explicit intermediate reasoning. CoT-VLA instead generates visual subgoals before actions and combines this reasoning with action chunking and hybrid attention.

  • Existing VLAs typically map language instructions and visual observations directly to robot actions without explicit intermediate reasoning steps.
  • CoT-VLA generates a subgoal image representing a planned future state before conditioning action generation on the observation and visual goal.The subgoal image serves as an intermediate visual reasoning step in pixel space.
  • Action-less video data can train subgoal image generation, expanding the data sources available for visual reasoning beyond action-annotated robot demonstrations.
  • CoT-VLA combines causal attention for text and image generation with full attention for action prediction, and incorporates action chunking.The model is built on a unified multimodal foundation model and trained on robot demonstrations plus action-less videos.
  • Experiments in simulation and the real world report improved policy performance over prior VLA approaches and state-of-the-art results across multiple platforms and tasks.

2. Related Work

Related work spans multimodal chain-of-thought reasoning and vision-language models used in robot perception, planning, and control. CoT-VLA connects these directions by using visual intermediate goals within an end-to-end VLA.

  • Multimodal chain-of-thought methods extend sequential reasoning beyond language by iteratively processing visual information.
  • Vision-language models have been used for task decomposition, object detection, dense rewards, goal generation, and pretrained representations in robot systems.
  • Recent VLA approaches fine-tune pretrained vision-language models on robot demonstrations for direct action prediction.

3. CoT-VLA

CoT-VLA predicts a future subgoal image and then an action sequence, training visual reasoning on both robot demonstrations and action-less videos. Its architecture uses VILA-U with hybrid attention and joint visual-action objectives.

  • 3.1. Visual Chain-of-Thought Reasoning: CoT-VLA uses robot demonstrations and action-less videos for pretraining, with language instructions and image sequences as inputs or targets.Robot demonstrations include actions, whereas action-less videos contain language descriptions and images without action annotations.
  • 3.1. Visual Chain-of-Thought Reasoning: The model first predicts a subgoal image n frames ahead, then predicts m actions conditioned on the current observation, instruction, and subgoal.
  • 3.2. The Base Vision-Language Model: Causal attention handles text and image generation, while full attention predicts action dimensions in parallel.The training objective combines visual-token and action losses.
  • 3.2. The Base Vision-Language Model: VILA-U provides autoregressive understanding and generation of image, video, and language tokens through a unified multimodal foundation model.The model uses discrete visual tokens and is pretrained on interleaved multimodal pairs.
  • 3.3. Training Procedures: Pretraining uses Open X-Embodiment robot data plus EPIC-KITCHENS and Something-Something V2 videos, with 256 × 256 images and action chunks of size 10.
  • 3.3. Training Procedures: LIBERO results report average success rates and standard errors across Spatial, Object, Goal, and Long suites using 3 seeds and 500 episodes per suite.

4. Experiments

Experiments evaluate CoT-VLA across simulation and real-world manipulation settings, comparing it with established baselines and isolating the effects of action chunking, hybrid attention, visual chain-of-thought reasoning, and pretraining. CoT-VLA achieves strong performance across benchmarks, with its highest average performance on Franka-Tabletop and a 46.7% relative pretraining improvement.

  • Experimental Setup: Evaluations span LIBERO simulation, Bridge-V2 real-robot manipulation, and Franka-Tabletop adaptation with 10–150 demonstrations per testing scenario.LIBERO contains four suites of 10 tasks with 50 demonstrations per task; Bridge-V2 uses 45k language-annotated trajectories.
  • Benchmark Results: CoT-VLA achieves best or competitive performance on LIBERO and comparable or better performance across Bridge-V2’s visual, motion, semantic, and language generalization categories.LIBERO evaluations use 500 trials per task suite with 3 random seeds, while Bridge-V2 categories use 10 trials each.
  • Benchmark Results: CoT-VLA achieves the highest average performance on Franka-Tabletop, improving across both single-instruction and multi-instruction scenarios.OpenX-pretrained models adapt better to multi-instruction tasks, while Diffusion Policy performs best on some single-instruction tasks.
  • Ablation Study: Action sequence prediction consistently outperforms single-action prediction, hybrid attention further improves performance, and CoT-VLA achieves the best LIBERO ablation results.The ablations evaluate action chunking, hybrid attention, and visual chain-of-thought reasoning on LIBERO-Spatial and LIBERO-Goal.
  • Ablation Study: 46.7% relative improvement, from 53.7% to 78.8%, is achieved by CoT-VLA with OpenX and action-less video pretraining over direct fine-tuning of the base VILA-U model.The comparison is conducted on Franka-Tabletop demonstrations and is reported as improved downstream task adaptation.
  • Visual Reasoning: CoT-VLA still struggles with out-of-distribution subgoal generation, while scaling large video and image models is identified as a promising direction.This limitation concerns visual reasoning generalization rather than the reported in-distribution benchmark comparisons.

5. Conclusion, Limitations and Future Work

CoT-VLA introduces explicit intermediate visual goals for VLA reasoning and demonstrates strong robotic manipulation performance. Its main limitations are inference overhead, lower visual quality, and discontinuities from action chunking.

  • Conclusion: CoT-VLA introduces intermediate visual goals as explicit reasoning steps, using subgoal images rather than abstract representations such as bounding boxes or keypoints.The system is built upon VILA-U and targets diverse robotic manipulation tasks.
  • Limitations: Generating 256 image tokens before action tokens causes a 7× average slowdown with an action chunk size of 10.Action chunking and parallel decoding improve speed, but image generation remains the primary bottleneck.
  • Limitations: Autoregressive image generation produces lower visual quality than state-of-the-art diffusion-based models.The authors identify faster unified multimodal models as a possible improvement direction.
  • Limitations: Action chunking can introduce discontinuous actions between chunks and lacks high-frequency feedback during execution.The authors suggest temporal smoothing and per-step prediction as potential remedies.

6. Implementation Details

The implementation details specify datasets, manually selected subgoal-horizon bounds, hyperparameter tables, fine-tuning settings, and substantial computational requirements.

  • Datasets: Open X-Embodiment provides robot demonstration pre-training data, while Something2Something and EPIC-KITCHEN-100 provide action-less video data.The predicted subgoal horizon uses manually set upper and lower bounds for each dataset.
  • Hyperparameters: Tables 4 and 5 present dataset weights and hyperparameters for pre-training.These tables summarize implementation settings used in the training pipeline.
  • Hyperparameters: The paper separately lists important hyperparameters for the model pre-training and pose-training stages.The implementation section frames these settings as key reproducibility details.
  • Fine-tuning: Fine-tuning on LIBERO and Franka-Tabletop uses a constant learning rate of 1e-5 for 150 epochs.The LLM backbone, projector, and depth transformer are fine-tuned.
  • Compute: Pre-training uses 12 A100 GPU nodes with 8 GPUs each and consumes 11K A100 GPU hours in total.LIBERO and Franka-Tabletop fine-tuning takes 10–24 hours on a single A100 GPU node, depending on dataset size.
Loading 2503.22020v1…