Trustworthy Evaluation of Robotic Manipulation: A New Benchmark and AutoEval Methods

Driven by the rapid evolution of Vision-Action and Vision-Language-Action models, imitation learning has significantly advanced robotic manipulation capabilities. However, evaluation methodologies have lagged behind, hindering the establishment of Trustworthy Evaluation for these behaviors. Current paradigms rely on binary success rates, failing to address the critical dimensions of trust: Source Authenticity (i.e., distinguishing genuine policy behaviors from human teleoperation) and Execution Quality (e.g., smoothness and safety). To bridge these gaps, we propose a solution that combines the Eval-Actions benchmark and the AutoEval architecture. First, we construct the Eval-Actions benchmark to support trustworthiness analysis. Distinct from existing datasets restricted to successful human demonstrations, Eval-Actions integrates VA and VLA policy execution trajectories alongside human teleoperation data, explicitly including failure scenarios. This dataset is structured around three core supervision signals: Expert Grading (EG), Rank-Guided preferences (RG), and Chain-of-Thought (CoT). Building on this, we propose the AutoEval architecture: AutoEval leverages Spatio-Temporal Aggregation for semantic assessment, augmented by an auxiliary Kinematic Calibration Signal to refine motion smoothness; AutoEval Plus (AutoEval-P) incorporates the Group Relative Policy Optimization (GRPO) paradigm to enhance logical reasoning capabilities. Experiments show AutoEval achieves Spearman's Rank Correlation Coefficients (SRCC) of 0.81 and 0.84 under the EG and RG protocols, respectively. Crucially, the framework possesses robust source discrimination capabilities, distinguishing between policy-generated and teleoperated videos with 99.6% accuracy, thereby establishing a rigorous standard for trustworthy robotic evaluation. Our project and code are available at https://term-bench.github.io/.

GPA-RAM: Grasp-Pretraining Augmented Robotic Attention Mamba for Spatial Task Learning

Most existing robot manipulation methods prioritize task learning by enhancing perception through complex deep network architectures. However, they face challenges in real-time collision-free planning. Hence, Robotic Attention Mamba (RAM) is designed for refined planning. Specifically, by integrating Mamba and parallel single-view attention, RAM aligns multi-view vision and task-related language features, ensuring efficient fine-grained task planning with linear complexity and robust real-time performance. Nevertheless, it has the potential for further improvement in high-precision grasping and manipulation. Thus, Grasp-Pretraining Augmentation (GPA) is devised, with a grasp pose feature extractor pretrained utilizing object grasp poses directly inherited from whole-task demonstrations. Subsequently, the extracted grasp features are fused with the spatially aligned planning features from RAM through attention-based Pre-trained Location Fusion, preserving high-resolution grasping cues overshadowed by an overemphasis on global planning. To summarize, we propose Grasp-Pretraining Augmented Robotic Attention Mamba (GPA-RAM), dividing spatial task learning into RAM for planning skill learning and GPA for grasping skill learning. GPA-RAM demonstrates superior performance across three robot systems with distinct camera configurations in simulation and the real world. Compared with previous state-of-the-art methods, it improves the absolute success rate by 8.2% (from 79.3% to 87.5%) on the RLBench multi-task benchmark and 40\% (from 16% to 56%), 12% (from 86% to 98%) on the ALOHA bimanual manipulation tasks, while delivering notably faster inference. Furthermore, experimental results demonstrate that both RAM and GPA enhance task learning, with GPA proving robust to different architectures of pretrained grasp pose feature extractors. The website is: https://logssim.github.io/GPA\_RAM\_website/.