Recursive Belief Vision Language Action Models
RB-VLA model excels in long-horizon tasks, achieving a 52.5% success rate improvement.
Key Findings
Methodology
RB-VLA is a belief-centric architecture trained with self-supervised world-model objectives, maintaining a compact latent state encoding task-relevant history, dynamics, and object interactions. Queried once per task, the VLM provides high-level intent, while the belief tracks task progress, enabling phase-aware, causally grounded control under partial observability.
Key Results
- RB-VLA achieves 52.5% and 37.5% higher success rates on multi-stage pick-and-place and stacking tasks, respectively, compared to π0, significantly reducing inference latency and eliminating memory growth observed in existing VLAs.
- In long-horizon benchmarks, RB-VLA outperforms prior VLAs, achieving 52.5% higher success rates on multi-stage pick-and-place tasks.
- Ablations show the belief module is the primary driver of performance, increasing success rates from 32.5% without belief to 77.5% with belief.
Significance
RB-VLA addresses inference latency and task progress loss issues in existing vision-language-action models for long-horizon tasks by introducing belief state representations. This model excels in multi-stage control tasks, holding significant academic and industrial implications.
Technical Contribution
RB-VLA combines belief states with diffusion policies for robust closed-loop execution under partial observability. Compared to existing methods, RB-VLA demonstrates higher success rates and lower inference latency in long-horizon tasks.
Novelty
RB-VLA is the first to integrate belief states with vision-language models, providing robust causal control in long-horizon tasks. It significantly innovates in task progress tracking and phase-aware control compared to existing methods.
Limitations
- RB-VLA may perform poorly under extreme occlusion or perceptual noise as the belief state relies on the accuracy of historical information.
- The model requires substantial simulated data for training, potentially limiting its applicability in data-scarce environments.
Future Work
Future work could focus on enhancing RB-VLA's robustness under extreme occlusion and reducing training data requirements to improve real-world applicability.
AI Executive Summary
The RB-VLA model aims to solve the limitations of existing vision-language-action models in long-horizon tasks, such as inference latency and task progress loss. By introducing belief states, RB-VLA achieves robust causal control under partial observability. Experimental results show that RB-VLA excels in multi-stage pick-and-place and stacking tasks, significantly improving success rates and reducing inference latency.
The key innovation of RB-VLA lies in its belief-centric architecture, trained with self-supervised world-model objectives to maintain a compact latent state encoding task-relevant history, dynamics, and object interactions. This model excels in long-horizon tasks, holding significant academic and industrial implications.
Despite RB-VLA's excellent performance in long-horizon tasks, it may perform poorly under extreme occlusion or perceptual noise. Future work could focus on enhancing the model's robustness and reducing training data requirements to improve its real-world applicability.
Deep Analysis
Background
Vision-language-action models have shown strong generalization by combining language understanding with visual perception to directly produce robot actions. However, most existing methods implicitly adopt a Markovian assumption, conditioning actions only on the current observation or a short history window. Such reactive reasoning fails to capture long-term context under noisy observations, perceptual aliasing, occlusion, and temporally static short observation histories.
Core Problem
Existing vision-language-action models face issues of inference latency and task progress loss in long-horizon tasks. Due to the lack of persistent, action-conditioned state representations, these models perform poorly in multi-stage control tasks.
Innovation
RB-VLA addresses inference latency and task progress loss issues in existing vision-language-action models for long-horizon tasks by introducing belief state representations. This model is trained with self-supervised world-model objectives, maintaining a compact latent state encoding task-relevant history, dynamics, and object interactions.
Methodology
- �� Belief-centric architecture: Trained with self-supervised world-model objectives, maintaining a compact latent state encoding task-relevant history, dynamics, and object interactions.
- �� Phase-aware control: Belief states provide short and long-horizon predictive context, supporting implicit phase estimation and causal control.
- �� Decoupled semantic grounding and control: Semantic reasoning is treated as episodic by querying a VLM once to extract grounded task intent, eliminating dense semantic re-inference.
Experiments
Experiments are conducted in RoboSuite and LIBERO-Long environments, evaluating RB-VLA's performance in long-horizon, multi-stage manipulation tasks. Training uses 40,000 simulated manipulation trajectories, assessing task success rate, inference latency, and memory growth.
Results
RB-VLA achieves 52.5% and 37.5% higher success rates on multi-stage pick-and-place and stacking tasks, respectively, compared to π0, significantly reducing inference latency and eliminating memory growth observed in existing VLAs.
Applications
RB-VLA is suitable for robotic applications requiring long-horizon task execution, such as industrial automation and home service robots. Its robust causal control capabilities make it excel in complex environments.
Limitations & Outlook
RB-VLA may perform poorly under extreme occlusion or perceptual noise as the belief state relies on the accuracy of historical information. The model requires substantial simulated data for training, potentially limiting its applicability in data-scarce environments.
Plain Language Accessible to non-experts
Imagine a robot working in a kitchen. It needs to find and pick up specific utensils from different cabinets. Traditional models are like a robot that only remembers the current task and has to check the instructions every time. RB-VLA is like a robot with memory, remembering previous task progress without needing to check the instructions each time. Even if a utensil is blocked by other items, it can continue the task based on previous memory.
ELI14 Explained like you're 14
Imagine you're playing a game where your task is to find and collect all the treasures in a room. Traditional robots are like players who have to check the map every time, while RB-VLA is like a player who remembers which rooms they've explored before. Even if a treasure is blocked by furniture, it can continue searching based on previous memory without checking the map every time. Isn't that cool?
Glossary
Belief State
A latent state representing task progress and dynamics, aiding in causal control under partial observability.
Used in RB-VLA to track task progress and enable robust control.
Vision-Language Model
A model that combines visual and language information for task reasoning.
Used in RB-VLA to extract task intent.
Self-Supervised Learning
A data-driven learning method without manual labeling.
Used to train RB-VLA's belief-centric architecture.
Diffusion Policy
A policy for generating continuous actions, combining belief states and task intent.
Used in RB-VLA for closed-loop control.
Multi-Stage Task
A complex task requiring multiple steps and stages to complete.
RB-VLA excels in multi-stage tasks.
Open Questions Unanswered questions from this research
- 1 How to enhance RB-VLA's robustness under extreme occlusion? Existing methods may perform poorly in such conditions, requiring improvements in belief state accuracy.
- 2 How to reduce RB-VLA's training data requirements? The current model requires substantial simulated data, limiting its applicability in data-scarce environments.
Applications
Immediate Applications
Industrial Automation
RB-VLA can be used in industrial robots for long-horizon task execution in complex environments, improving production efficiency.
Long-term Vision
Home Service Robots
RB-VLA can be used in home service robots, helping them execute multi-stage tasks in complex home environments, such as tidying rooms or preparing food.
Abstract
Vision-language-action models must enable agents to execute long-horizon tasks under partial observability. However, most existing approaches remain observation-driven, relying on short context windows or repeated queries to vision-language models (VLMs). This leads to loss of task progress, action repetition under perceptual aliasing, and high inference latency. While semantic grounding is important, long-horizon manipulation fundamentally requires persistent, action-conditioned state representations. Current VLAs lack such representations and exhibit limited temporal and physical reasoning, making them ill-suited for multi-stage control. This paper introduces RB-VLA, a belief-centric architecture trained with self-supervised world-model objectives that maintains a compact latent state encoding task-relevant history, dynamics, and object interactions. Queried once per task, the VLM provides high-level intent, while the belief tracks task progress and enables phase-aware, causally grounded control under partial observability without storing raw observations or scaling memory with time. The belief and intent jointly condition a diffusion policy for robust closed-loop execution. RB-VLA outperforms prior VLAs on long-horizon benchmarks, achieving 52.5 percent and 37.5 percent higher success rates on multi-stage pick-and-place and stacking tasks, respectively, compared to pi_0. It also reduces inference latency by up to five times relative to baselines and eliminates memory growth across timesteps observed in existing VLAs. Ablations show the belief module is the primary driver of performance, increasing success rates from 32.5 percent without belief to 77.5 percent with belief.