Reliable real-world deployment of large-scale VLA policies

Establish reliable methods for deploying large-scale vision-language-action policies, including OpenVLA, pi_0, MolmoAct, and Gemini Robotics, in real-world robotic settings.

Background

The paper notes that while large-scale vision-language-action (VLA) policies such as OpenVLA, pi_0, MolmoAct, and Gemini Robotics have demonstrated strong language-conditioned manipulation capabilities across multiple embodiments, consistently reliable deployment outside lab settings has not been achieved.

This open problem motivates the work’s focus on scalable, generalizable reward modeling—specifically, leveraging zero-shot progress estimation from pretrained video VLMs—to enable reinforcement learning and policy improvement without brittle, hand-crafted rewards.

References

Large-scale vision-language-action policies such as OpenVLA, pi_0, MolmoAct and Gemini Robotics have demonstrated strong language-conditioned manipulation capabilities across diverse embodiments, yet reliably deploying them in real-world settings remains an open problem.

— TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics  (2602.19313 - Chen et al., 22 Feb 2026) in Related Work — The reward bottleneck for VLA.

Several claims remain deliberately open. The current snapshot does not provide a uniform upstream-versus-FluxVLA reproduction study, a controlled comparison of all LIBERO models, multi-seed uncertainty for most table entries, a single-/multi-node training scaling sweep, task-level parity for accelerated inference, a controlled comparison of RTC routes under injected latency, a local-versus-remote closed-loop comparison, or a common multi-robot trial benchmark.

— FluxVLA Engine: A One-Stop VLA Engineering Platform for Embodied Intelligence  (2609.17210 - Li et al., 15 Sep 2026) in Section 6, subsection “Findings and remaining controlled studies”

All quantitative experiments use a single model family ($\pi_{0.x}$) in simulation, leaving open whether the signal transfers across VLA architectures and to real-world deployment.

— Decoding Task Progress from VLA Representations  (2608.13474 - Bhardwaj et al., 13 Aug 2026) in Discussion and Limitations, Section 6

The retained softmax blocks still impose quadratic cost, and transfer to video and real-world robotic deployment remains to be established.

— IronViT: Toward Efficient Generalist Visual Representation Learning  (2609.29252 - Huang et al., 24 Sep 2026) in Section Conclusion

Future work should test whether pre-execution signals can predict when correction helps and validate the measured forward-time gains in a real-robot control loop.

— Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies  (2609.21216 - Tang et al., 18 Sep 2026) in Section Discussion And Limitations

Real-robot validation remains future work, though the CPU latency and the SO-101-class hardware support in the underlying framework make it direct.

— MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?  (2609.03715 - Sendai et al., 3 Sep 2026) in Section DISCUSSION AND LIMITATIONS, subsection “Limitations”

Experimental results indicate that language-conditioned UAV control remains a challenging open problem, highlighting substantial opportunities for future advances in multimodal representation learning, continuous action prediction and cross-embodiment transfer.

— WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses  (2610.08526 - Le et al., 6 Oct 2026) in Section 1, Introduction

The actionable recommendation is therefore nuanced: under a tight compute budget SmolVLA is the rational default, but when held-out fidelity matters $\pi_0$ is preferable; critically, neither reaches deployable accuracy on this benchmark, which remains open.

— WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses  (2610.08526 - Le et al., 6 Oct 2026) in Section 5.1, Deployment-oriented analysis, subsection Cost--accuracy frontier

Deploying DaCe-DT in open-world or real-robot settings, where observations are noisy, tasks are ambiguous, and reward feedback is sparse or absent, remains an important open challenge.

— DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks  (2610.11085 - Wang et al., 8 Oct 2026) in Section "Limitations"