RB-VLA: Recursive Belief Vision-Language Model

• RB-VLA is a recursive belief vision-language-action model that maintains a persistent, action-conditioned internal state to guide long-horizon manipulation under partial observability.
• It decouples high-level intent extraction from real-time belief state estimation, using a one-time VLM query and a diffusion policy to generate coherent action sequences.
• Empirical evaluations show significant success improvements in tasks like pick-and-place and stacking, demonstrating reduced inference latency and efficient memory usage.

Searching arXiv for the cited RB-VLA papers and closely related entries to ground the article. {"query":"RB-VLA Recursive Belief Vision LLM arXiv (Bagaria et al., 24 Feb 2026)"} {"query":"RouterVLA arXiv (Ren et al., 25 Jun 2026)"} RB-VLA, short for Recursive Belief Vision–Language–Action Model, is a belief-centric architecture for long-horizon robot manipulation in realistic, partially observable settings. It is designed for the regime in which current vision-language-action models remain observation-driven, condition actions on the current image or a short history window, and often re-query a large vision-LLM at every step. RB-VLA instead maintains a persistent, action-conditioned internal state, queries a vision-LLM once per episode to obtain a high-level intent, and conditions a diffusion policy on the evolving belief and fixed intent for phase-aware closed-loop control under partial observability (Bagaria et al., 24 Feb 2026).

1. Problem setting and motivating failure modes

RB-VLA targets long-horizon, multi-stage manipulation under partial observability. In this setting, the central claim is not that semantic reasoning alone is insufficient in a generic sense, but that semantic reasoning alone is not the primary bottleneck in long-horizon manipulation. The reported failure modes of existing VLA pipelines are instead tied to the absence of persistent, action-conditioned state representations and to limited temporal and physical reasoning.

The architecture is motivated by four recurrent deficiencies in predominantly observation-driven pipelines. First, perceptual aliasing and action repetition arise when multiple scenes or objects look similar but correspond to different phases of the task. Second, forgetting past progress occurs when short-term context causes policies to re-grasp an already placed object or otherwise lose track of task stage. Third, high inference latency follows from repeatedly running heavyweight VLMs. Fourth, memory growth with horizon appears when context is extended by accumulating visual or memory tokens.

RB-VLA addresses these issues by replacing stepwise semantic re-interpretation with a persistent belief that captures task-relevant history and dynamics. The high-level intent is extracted once at the beginning of an episode and kept fixed thereafter. This design is intended to preserve task progress, disambiguate visually similar scenes at different phases, and eliminate the memory growth associated with accumulating raw observations across time.

2. POMDP formulation and recursive belief dynamics

The model formalizes long-horizon manipulation as a partially observable Markov decision process with tuple $(S, A, O, T, R)$, where $S$ is the latent world state, $A$ the continuous action space, $O$ the visual and proprioceptive observations, $T$ the transition dynamics $p(s_{t+1} \mid s_t, a_t)$, and $R$ an optional reward. RB-VLA is trained by self-supervised world-model objectives and behavior cloning rather than reinforcement learning.

Because $o_t$ is partial and noisy, the architecture maintains a latent belief $z_t$ that summarizes history and enables filtering. The belief update is written as

$z_t = f_\theta(z_{t-1}, a_{t-1}, o_t),$

with probabilistic factorization

$S$0

and an observation model

$S$1

In implementation, the belief module is an RSSM-style world model with a deterministic recurrent state $S$2 and a stochastic latent $S$3. Its generative components are a prior over the stochastic latent,

$S$4

an inference network,

$S$5

where $S$6 is the evidence token extracted from the current observations conditioned on $S$7, and a deterministic update

$S$8

The observation or feature prediction term is

$S$9

where $A$0 is a visual/proprioceptive latent target from an EMA teacher.

This construction makes the belief both persistent and action-conditioned. The paper reports that it captures contact events, object interactions, and progress even under occlusion, while remaining compact and not growing with time. A plausible implication is that the model treats latent state estimation, rather than longer and longer observation context, as the primary mechanism for long-horizon control.

3. Architectural decomposition: intent, belief, and diffusion control

RB-VLA decouples semantic grounding from control through three modules: a vision-language module queried once per episode, a belief estimator that recursively compresses interaction history and action-conditioned dynamics into a fixed-size latent state, and a diffusion policy that produces action chunks conditioned on belief and intent.

The VLM is Qwen2.5-VL-7B (frozen). At $A$1, the instruction and initial image are passed through the VLM, whose final multimodal hidden states

$A$2

are pooled into a single task vector via a learned goal query,

$A$3

followed by an MLP that yields a compact intent embedding

$A$4

The intent remains fixed throughout execution, on the assumption that semantic goals and object identities do not change within an episode.

The belief estimator uses a DINOv2 ViT-S/14 backbone to encode visual observations into patch tokens $A$5. A learned query $A$6 computes attention

$A$7

and the frame summary

$A$8

emphasizes scene regions predictive of robot–object dynamics. A temporal transformer integrates a fixed window of $A$9 frame summaries together with the previous belief token $O$0, prepended to the sequence, yielding the evidence token $O$1.

The low-level controller is a diffusion transformer policy trained to model

$O$2

where $O$3 is a fused state encoder combining proprioception and DINOv2 features to a shared dimension. It uses a standard conditional DDPM objective with 100 diffusion steps to predict 16-step action chunks:

$O$4

At test time, iterative denoising yields a chunk; the first action is executed at 50 Hz, and the remainder can warm-start the next sample with low-variance noise to improve temporal coherence. The resulting policy is summarized as

$O$5

The closed-loop execution path is fixed-memory. At $O$6, the frozen VLM is queried once to obtain intent. For subsequent timesteps, the model encodes the current observation and proprioception, forms evidence with the temporal transformer, samples or takes the mean of $O$7 from the prior at inference, updates the belief with the GRU, and conditions the diffusion policy on $O$8 to sample an action chunk. Only the 256-dimensional belief persists across timesteps.

4. Training objectives and implementation regime

The belief module is trained with a world-model ELBO and an inverse-dynamics consistency loss. The world-model loss is

$O$9

The auxiliary 5-step prediction head is included to extend temporal foresight. Action-grounding is enforced by

$T$0

The training corpus comprises 40,000 simulated trajectories (~40M timesteps) across 10 tasks in RoboSuite, LIBERO-Long, LIBERO-Object, and a UR5 MuJoCo environment. Visual augmentations are brightness/contrast jitter, Gaussian noise, and mild crops. The system is further fine-tuned on 100 real-world UR5 trajectories to adapt to sensor noise and unmodeled dynamics.

The belief module has ~150M parameters, temporal context $T$1, and belief dimensionality 256; it is trained for 80k iterations with batch size 128 and learning rate $T$2. Target embeddings for $T$3 are derived from an EMA teacher. The scheduling of $T$4 and $T$5 is reported to encourage exploration early and sharpen the belief later. The stochastic latent is included to avoid mode collapse and maintain multimodal hypotheses.

The intent module uses the frozen Qwen2.5-VL-7B backbone and a learned attention pool followed by an MLP to 256 dimensions. It is trained end-to-end with the diffusion policy via action prediction loss, but the VLM itself is not re-queried during an episode. The diffusion policy has ~120M parameters, is trained for 100k iterations with batch size 32 and learning rate $T$6, and runs in receding-horizon mode at 50 Hz. Training uses 12× NVIDIA A100 80GB, while inference measurements are taken on a single A100 (Bagaria et al., 24 Feb 2026).

5. Empirical performance, latency, and representation analysis

RB-VLA is evaluated on long-horizon, multi-stage manipulation in RoboSuite and LIBERO-Long, with 40 trials per task under randomized initializations, object identities and placements, lighting, dropped frames, and observation noise. Success is defined as task completion within the horizon without collisions or early termination.

On pick-and-place under partial observability, RB-VLA achieves 82.5% on the single-object setting, compared with 55.0 for $T$7. In the multi-stage setting, GROOT-N1 attains 20.0, OpenVLA attains 22.5 and is described as failing on multi-stage in this setting, $T$8 attains 25.0, and RB-VLA attains 77.5. The reported improvements are 52.5 percentage points over $T$9 and 57.5 points over GROOT-N1. On stacking, RB-VLA reaches 85.0 on single-stack versus 67.5 for $p(s_{t+1} \mid s_t, a_t)$0, and 80.0 on multi-stack versus 42.5 for $p(s_{t+1} \mid s_t, a_t)$1, corresponding to a 37.5-point improvement.

Latency and memory behavior are central to the evaluation. Because the VLM is queried only once and history tokens are not accumulated, RB-VLA reduces inference latency by up to 5× relative to baselines that re-invoke large VLMs or process multi-frame contexts. Relative GPU memory usage remains at 1.0× from 1 to 8 frames, whereas Qwen2-VL-7B grows from 1.00× to 6.82× and GR00T-N1 from 1.00× to 7.12× over the same context lengths.

The belief representation is analyzed directly. Similarity in belief space correlates with similarity of future actions at Pearson $p(s_{t+1} \mid s_t, a_t)$2 and future visual embeddings at $p(s_{t+1} \mid s_t, a_t)$3, which is presented as evidence that $p(s_{t+1} \mid s_t, a_t)$4 functions as a latent dynamical state. Sampling $p(s_{t+1} \mid s_t, a_t)$5 from the prior produces diverse yet structured futures, with a non-zero stabilized KL indicating sustained stochastic usage. Attention visualizations further show that, under occlusion, the temporal transformer attends to the recursive belief token and causally relevant past frames.

The ablation study on two-object pick-and-place under partial observability reports a graded increase in success: 32.5% with no frame-encoder targets, no stochastic $p(s_{t+1} \mid s_t, a_t)$6, and no belief-to-policy conditioning; 57.5% when belief-to-policy conditioning is added; 62.5% when stochastic $p(s_{t+1} \mid s_t, a_t)$7 is also added; and 77.5% for full RB-VLA. This is summarized as evidence that the belief module is the primary performance driver, increasing success from 32.5% to 77.5%. In real-world deployment, on a UR5 with an Intel RealSense D435i and only RGB input, the model achieves 68.0% success over 25 trials on multi-object pick-and-place under occlusion after fine-tuning with 100 real trajectories (Bagaria et al., 24 Feb 2026).

6. Limitations, implications, and acronym disambiguation

The reported advantages of RB-VLA are explicitly empirical rather than theoretical guarantees. The model may still struggle under extreme or prolonged occlusions that invalidate predictive targets, under severe distribution shifts that break the visual encoder or belief dynamics, or when the single VLM query yields incorrect or underspecified high-level intent. Suggested future directions are scaling data/task diversity, integrating reinforcement learning for recovery behaviors, and extending belief adaptation for more dynamic environments.

Within the scope of the reported experiments, the practical implication is that explicit belief-state estimation, learned with self-supervised world-model objectives and coupled with a single-shot intent from a VLM, can improve robustness, efficiency, and scalability of long-horizon, multi-stage manipulation under partial observability. This suggests a system design in which semantic grounding and low-level control are separated, with persistent latent state carrying the burden of phase tracking, occlusion handling, and action conditioning.

A common source of confusion is that RB-VLA is not a unique acronym. In “RouterVLA: Turning Smoke Tests into Supervision for Heterogeneous VLA Selection”, RB-VLA refers instead to a router-/rule-based Vision-Language-Action selection layer that reuses pre-deployment smoke tests to choose among frozen, heterogeneous VLA experts per task and perturbation condition (Ren et al., 25 Jun 2026). That usage concerns commissioning-aware policy selection over an expert pool, whereas Recursive Belief Vision–Language–Action Model denotes a single belief-driven control architecture. The shared acronym therefore identifies two distinct research programs: one centered on latent-state control for partially observable manipulation, and one centered on routing among heterogeneous VLAs.