Sep 2026· ITU Journal on Future and Evolving Technologies· 0 citations
TL;DR
The results show that value models provide useful trajectory ranking and failure-detection signals in the studied manipulation setting and improve assisted task completion and reduce takeover frequency relative to recovery-only training, although these gains must be distinguished from fully autonomous policy performance.
Abstract
Vision Language Action (VLA) models have become an important approach to robotic manipulation by integrating visual observations, language instructions, and robot states for action generation. Nevertheless, behavior cloning-based VLA policies remain vulnerable to distribution shifts, accumulated execution errors, and failure recovery difficulties during long horizon closed-loop operation. This paper investigates whether an auxiliary value architecture can improve the reliability of VLA policy learning without replacing the policy as the low-level controller. We present a controlled real-robot study on a LeRobot SO-101 platform covering block grasping, placing, and sorting tasks. The study compares three policy architectures, namely ACT, diffusion policy, and SmolVLA, and two value architectures, namely ValueLite and ValueVision. The value models estimate state quality, identify early failure precursors, and generate Advantage-Conditioned Policy (ACP) labels for policy retraining. Our evaluation considers policy-only success, human-assisted success, intervention rate, recovery success, completion time, action smoothness, value prediction error, early failure detection, and inference cost. In addition, data-controlled and mechanism-level ablations are designed to separate the effects of additional recovery or rollout data from the effects of learned value signals and ACP conditioning. The results show that value models provide useful trajectory ranking and failure-detection signals in the studied manipulation setting. Value-guided variants improve assisted task completion and reduce takeover frequency relative to recovery-only training, although these gains must be distinguished from fully autonomous policy performance.
Robots operating in open environments act under partial observability, physical constraints, and dynamic task contexts. Beyond mapping observations and language instructions to actions, they must anticipate how candidate actions may affect future states and task-relevant outcomes. Recent advances in world models, video...
Zu-Xing Lu, Hong-Jia Zhai, Guan-Zhi Wang et al.· arXiv.org· 6 citations
Vision-language-action (VLA) intelligence provides a unified route for connecting visual perception, language understanding, and executable control in traditional industrial robotic arms. This paper examines the transition from modular perception-planning-control pipelines to generalist policies that predict robot acti...
Wei-Jian Wan· Applied and Computational En...· 0 citations
Within this trained-task regime, DEM sits on the observed success--latency--energy frontier and provides a strong, efficient baseline for language-conditioned robot skills.
Results show that directly steering internal robot-policy representations can improve the performance of robot policies at inference time and show that directly steering internal robot-policy representations can improve the performance of robot policies at inference time.
Jiu-Zhou Lei, Chang Liu, Da-You Li et al.· 0 citations
ActionGround is presented, a neuro-symbolic, training-free runtime layer that wraps a frozen VLA policy without retraining, fine-tuning, or weight access, adding less than 1 ms of overhead per control step.
N. Chandra, Madhur Thareja, Shriram Damodaran et al.· arXiv.org· 0 citations
Vision-Language-Action (VLA) models have emerged as a powerful paradigm for embodied intelligence, but fine-tuning them with reinforcement learning (RL) remains constrained by the cost of real-world robot interaction. Model-based reinforcement learning (MBRL) reduces this cost by using a learned world model to generate...