VLA Research Advances: 2026
Sources: arXiv weekly digests, 2026-W07 through 2026-W32; Moritz Reuss, 2025-10-01; Physical Intelligence, pi.website/research/rlt, 2026-03; Zhang et al. (CoRL 2025), rewind-reward.github.io, 2025-10 Raw: W07; W08; W09; W10; W11; W12; W13; W14; W15; W16; W17; W18; W19; W20; W21; W22; W23; W24; W25; W26; W27; W28; W29; W30; W31; W32; [ICLR 2026 VLA State](../../raw/clippings/State of VLA Research at ICLR 2026.md); RLT; ReWiND Updated: 2026-08-12
Overview
The 2026 arXiv landscape (cs.RO/cs.LG/cs.HC/cs.CV, W07–W32) shows five dominant research directions in VLA and robot learning: test-time compute scaling for VLAs, identifying information bottlenecks that limit VLA scaling, VLA safety (both attack vectors and defenses), foundation models for tactile manipulation, and memory/state augmentation for long-horizon tasks. The volume of VLA papers approximately doubled quarter-over-quarter from W07 to W25.
Test-Time Compute Scaling
RD-VLA (arXiv:2602.07845, Recurrent-Depth VLA): addresses the inefficiency of fixed-compute VLAs by adding a recurrent, weight-tied action head that performs latent iterative refinement at inference time. Key architectural properties:
- Training: truncated backpropagation through time (TBPTT)
- Inference: adaptive stopping criterion based on latent convergence — more compute allocated only when task demands it
- Memory: does not increase memory relative to single-pass inference
Results: tasks with 0% success at single-iteration inference reach >90% success at four iterations; up to 80× speedup over prior reasoning-based VLA models that use Chain-of-Thought token generation.
Significance: demonstrates that VLA test-time compute can be scaled without the memory penalty of Chain-of-Thought. The latent iterative refinement paradigm is distinct from both token-level scaling and diffusion denoising — it refines in the action head’s latent space.
The Compression Gap: VLA Scaling Bottlenecks
Compression Gap principle (arXiv:2604.03191): scaling behavior in VLA pipelines is governed by the tightest information bottleneck, not by the largest component. Two action representation regimes:
- Continuous actions (Diffusion Policy): vision encoder is the binding constraint. Upgrading the encoder improves performance by +21 percentage points on LIBERO.
- Discretized actions (OAT, fixed-capacity codebook): the codebook becomes the bottleneck. Encoder upgrades are substantially attenuated — improvements are absorbed by the codebook’s limited capacity.
Implication: VLA scaling cannot be planned around a single component. The pipeline must be analyzed for its tightest bottleneck, and that bottleneck must be addressed first. This is why diffusion-based VLAs (π₀, SmolVLA) have outpaced autoregressive token-based VLAs at equivalent parameter counts.
Action Hallucination
Action hallucination (arXiv:2602.06339) is a named failure mode in generative VLAs: the model generates plausible-looking action sequences that are physically inconsistent with the observed scene. Distinct from standard hallucination in language models — the output format (continuous joint commands) looks correct but conflicts with the actual state of the environment. Emerging as a systematic evaluation dimension alongside task success rate.
Memory and Long-Horizon State
Notes-to-Self (arXiv:2602.21013): scratchpad-augmented VLAs that maintain explicit written notes across decision steps. The scratchpad operates in language space — the VLA writes to it and reads from it — enabling retrieval of information from many steps prior without relying on the context window.
ECHO (arXiv:2605.10993): continuous hierarchical memory specifically for VLAs. Addresses goal drift in long-horizon tasks: “Never do X” constraints disappear from context after enough decision steps.
EA-WM (arXiv:2606.13053): Event-Aware World Model. Augments frozen visual-feature dynamics with task-specification-grounded event prediction — decodes imagined futures into structured event states (object displacement, placement predicates) and scores them on task-progress, semantic-consistency, physical-feasibility, and uncertainty. Makes feature-space world models more interpretable for task progress assessment.
Inference Efficiency
Several 2026 papers target the inference latency bottleneck for real-time robot control:
- QuantVLA (arXiv:2602.20309): scale-calibrated post-training quantization for VLAs — maintains task performance at lower bit widths
- Latent Bridge (arXiv:2605.02739): feature delta prediction for dual-system VLAs; decouples slow VLM inference from fast action generation
- Real-Time Autoregressive Policies (arXiv:2606.13355): enabling token-by-token autoregressive generation at robot control frequencies
- NoRD (arXiv:2602.21172): data-efficient VLA that skips explicit reasoning steps — direct perception to action without chain-of-thought overhead
Online RL for VLA Adaptation
Two systems demonstrate that small RL components trained on top of frozen VLAs can dramatically accelerate task-specific performance without retraining the base model.
RLT — RL Tokens (Physical Intelligence, March 2026): extracts a compact “RL token” from VLA embeddings via an encoder-decoder bottleneck; trains a small actor-critic directly on this compressed state representation.
- Actor receives the VLA’s predicted action as input, so it “learns to edit the VLA action rather than replace it entirely”
- Updates run at hundreds per second onboard
- Regularization keeps exploration near baseline behavior, deviating only when improvements are found
Precision task results (successes/10 min):
| Task | Base VLA | RLT |
|---|---|---|
| Screwdriver | 1.7 | 14 |
| Zip Tie | 2.8 | 13 |
| Ethernet | 147 | 400 |
| Charger | 136 | 600 |
50% of ethernet insertion trials exceeded all human demonstration speeds.
Significance: first deployed demonstration that a frozen VLA can be wrapped with a small RL component for online improvement — without modifying VLA weights. The RL token serves as the critic’s state representation, enabling sample-efficient off-policy RL in a compact latent space.
ReWiND (Zhang et al., CoRL 2025): language-conditioned reward model + offline pretraining + online RL for adaptation to unseen tasks without per-task demonstrations.
- Reward model trained on a small demo dataset + curated Open-X subset, augmented with LLM-generated instructions and video rewinding
- Pre-trains a language-conditioned policy via offline RL (IQL), then fine-tunes online with reward model labels
- Results: 79% success rate after 100k environment steps on 8 unseen tasks in Meta-World, 2× better than best baseline
- Real-world: 5× improvement on bimanual tabletop tasks in just 1 hour of online interactions
Key innovation — video rewinding: playing a successful demonstration in reverse generates a “failure” trajectory with naturally decreasing progress scores, providing dense reward signal for training even when the policy is making mistakes. This avoids the need for separate failure demonstration collection.
Foundation Tactile Policy
FTP-1 (arXiv:2606.13102, Generalist Foundation Tactile Policy): the first cross-sensor foundation model for tactile manipulation:
- Scale: pretrained on ~3,000 hours of data from 26 sources spanning 21 different tactile sensors
- Architecture: heterogeneous encoders project diverse tactile inputs (image-based, array-based, state-based) → unified morphology-aware latent tokens → shared tactile Transformer expert
- Results: +17.2% on seen sensor setups; +31% success rate gain on two previously unseen tactile sensor configurations
- Significance: analogous to foundation vision models but for tactile sensing; the cross-sensor generalization (+31% on unseen sensors) is the key capability — prior tactile policies were hardware-specific
AT-VLA (arXiv:2605.07308): adaptive tactile injection for VLAs — adds tactile signal conditioning to standard VLA architectures without requiring tactile sensors at training time. Enables using existing vision-only VLA architectures with tactile input during deployment.
TAMEn (arXiv:2604.07335): Tactile-Aware Manipulation Engine for closed-loop data collection in contact-rich tasks — addresses the data collection bottleneck for tactile policies.
VLA Safety: Attack Vectors
Trajectory-Level Redirection Attacks (arXiv:2606.12978): a stronger threat model than prior adversarial VLA work. Attacker specifies one text prompt before an episode; the prompt is semantically close to the benign instruction (omitting only target words); all policy and environment components remain fixed. An on-policy prompt search discovers closed-loop perturbations that redirect the robot’s entire physical trajectory to an attacker-specified target. Validated both in simulation and on hardware. Implication: VLA text prompts are a persistent trajectory-level control channel exploitable without modifying any model weights.
VLA-Forget (arXiv:2604.03956): machine unlearning for VLAs — removing specific behaviors (skills, safety violations, proprietary demonstrations) from a trained VLA without retraining from scratch. Relevant for responsible deployment and capability withdrawal.
Belief Dynamics for Safe Collaborative Manipulation (arXiv:2604.04967): detects behavioral shifts in human collaborators during manipulation (e.g., human switching from cooperative to non-cooperative intent) to trigger safety interventions.
VLA Safety: Defenses
Safe VLA via Barrier Enhanced Flow Matching (arXiv:2607.29569): embeds Control Barrier Function (CBF) constraints directly into the flow matching denoising process. Rather than filtering outputs post-hoc (which can be overly conservative), CBF constraints modify intermediate denoising steps to steer trajectories away from unsafe regions. Result: provably safe VLA inference while preserving the expressiveness of the underlying flow-matching model. Modular — can wrap any pretrained flow-matching VLA without retraining.
Bimanual and Humanoid
GR-Dexter (ByteDance Research, March 2026): VLA framework for bimanual dexterous control at 21-DOF.
- Hardware: ByteDexter V2 Hand (21 DOF, piezoresistive fingertip sensors), teleoperated via Meta Quest VR + Manus Metagloves + foot pedals
- Architecture: 4B Mixture-of-Transformer (MoT)
- Training: co-trained on web-scale vision-language data + cross-embodiment robot data (Fourier ActionNet, OpenLoong, RoboMIND) + human VR trajectories
- Results: 97% success in-distribution, 89% out-of-distribution (vs. 64% baseline)
- Key finding: cross-embodiment data significantly improves OOD generalization; VR trajectories with temporal consistency filtering are a viable bimanual training signal
CRAFT (arXiv:2604.03552): video diffusion for bimanual robot data generation — synthesizes bimanual demonstration video from single-arm demonstrations, expanding dataset diversity without additional hardware collection.
BifrostUMI (arXiv:2605.03452): robot-free demonstrations for humanoid whole-body manipulation, extending the UMI handheld gripper concept to whole-body tasks.
HEX (arXiv:2604.07993): Humanoid-Aligned Experts for Cross-Embodiment whole-body manipulation — transfers manipulation policies across humanoid embodiments.
Evaluation
LIBERO-X (arXiv:2602.06556): a robustness evaluation protocol for VLAs with three progressive difficulty tiers: spatial generalization, object recognition, task instruction understanding. Representative VLA models show significant performance drops under cumulative perturbations — exposing persistent limitations in scene comprehension and instruction grounding that standard success-rate metrics mask.
ICLR 2026 Research Landscape
Moritz Reuss’s practitioner survey (October 2025) documents the explosive growth of VLA research: ICLR 2024 received 1 VLA submission, ICLR 2025 received 9, and ICLR 2026 received 164 — an 18x year-over-year increase.
Nine research directions at ICLR 2026:
- Discrete Diffusion VLAs: parallel action sequence generation vs. sequential autoregressive; generates 100-step sequences in few forward passes. Combined with ECoT for simultaneous sub-goal and reasoning generation.
- Reasoning VLAs / ECoT: spatially-grounded predictions (bounding boxes, 2D trajectories) + subtask decomposition. Helps align representations between static VLM pretraining and embodied control.
- New Discrete Tokenizers: Residual Vector Quantization + frequency/time-domain losses, spline-based parameterization, DCT-inspired objectives for physically plausible motion.
- Efficient VLAs: smaller architectures, better tokenizers, quantization, distillation, hypernetworks. Task-specific policy generation conditioned on instructions + initial observations.
- RL for VLAs: residual RL with small correction policies on top of frozen VLAs; stage-aware RL decomposing tasks into semantic phases (reach→grasp→transport→place) with stage-level rewards.
- VLA + Video Prediction: initializing from video foundation models + future-frame and action prediction. Limitation: diffusion/flow-based video models face slow inference.
- Evaluation and Benchmarking: real-to-sim translation for automatic environment construction; world models as evaluation environments.
- Cross-Action-Space Learning: soft-prompting tokens for different datasets, unified vision-motion representations via shared codebooks, hierarchical mixture-of-experts for embodiment adaptation.
- Memory/Composition and Other: memory modules aggregating prior context; policy composition via score-summing (diffusion/flow formulations); VLM backbone selection uncorrelated with standard VLM benchmarks.
Two underrepresented problems identified in the survey: (1) data quality curation — few submissions address demonstration data quality despite acknowledged OXE data quality concerns; (2) in-context learning — language alone insufficient for complex physical tasks, but limited VLA work explores this despite LLM/VLM success.
Benchmark saturation guide (Reuss):
- LIBERO: ≥95% is the current standard; minimal discrimination between 95–98% results
- CALVIN ABC (cross-setup): >4.0 standard, >4.5 state-of-the-art
- CALVIN D (fine-tuning): >3.75 standard, >4.0 very good
- CALVIN ABCD (multi-setup): >4.5 relevant
- SIMPLER: highly variable (40–99% on Bridge alone); Google Robot ~70–80% for current SotA
The frontier vs. open-source gap: despite open-source VLAs matching frontier performance on simulation benchmarks, significant gaps emerge in zero-shot open-world behavior. Root causes: benchmark saturation, limited open-source data diversity, insufficient understanding of high-quality demonstration data, narrow evaluation scope.
Open Questions
- Does the Compression Gap principle apply equally to flow-matching VLAs vs. diffusion VLAs? Are there additional bottleneck types beyond vision encoder and action codebook?
- FTP-1’s +31% on unseen sensors was measured on two sensors. Does cross-sensor generalization hold for novel sensor modalities (e.g., distributed skin sensing) not represented in pretraining?
- Trajectory-level prompt attacks were validated on hardware — what defenses work beyond CBF (which addresses kinematic safety, not semantic redirection)?
- Can Notes-to-Self / ECHO solve goal drift in VLA loops without growing context window requirements?